Paper Title (use style: paper title) Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 772 Harnessing Machine Learning and Real-Time Object Detection to Revolutionise Industrial Quality Control Dr Vivek Uprit1, Dr Ravi Mohan2, Dr Bharti Bhattad3 1Associate Director, Dataflair Web Services Pvt Ltd., Indore, MP, India, Email ID: vivek.uprit@data-flair.training 2Associate Professor, Makhanlal Chaturvedi National University of Journalism and Communication, Bhopal, MP, India, Email ID: ravi@mcu.ac.in 3Associate Professor, Acropolis Institute of Technology and Research, Indore, MP, India, Email ID: bhartibhattad118@gmail.com Article History: Received:04/12/2024 Revised:03/01/2025 Accepted:05/02/2025 Abstract: This study introduces an innovative real-time object detection system designed to improve quality control processes within industrial manufacturing environments. Traditional inspection methods, often based on manual checks or periodic sampling, can be slow, labour-intensive, and prone to human error, leading to potential defects reaching consumers and increased rework costs. In contrast, the proposed system employs advanced computer vision techniques, leveraging convolutional neural networks (CNNs) and region proposal networks (RPNs) to automatically and efficiently identify and locate objects and defects as they appear on the production line. The system's key advantage lies in its ability to deliver rapid, precise detection with minimal latency, enabling immediate responses to quality issues. This real-time functionality helps prevent defective products from progressing further in the manufacturing process, thus reducing waste, rework, and associated costs. Additionally, the approach is highly adaptable, capable of accommodating different product types, sizes, and orientations, making it suitable for a wide range of industrial applications. Furthermore, the system incorporates dynamic scaling mechanisms that allow it to adjust seamlessly to variations in the items being inspected, ensuring consistent high standards across diverse manufacturing contexts. The study also highlights the system's flexibility in customisation, enabling it to meet specific industry requirements and conditions. Experimental results demonstrated that this method outperforms traditional inspection approaches, achieving higher accuracy and faster detection times, and offering a more sustainable and efficient solution for quality assurance. Overall, this research underscores the transformative potential of integrating advanced computer vision technologies into industrial quality control, setting a new standard for accuracy, speed, and operational efficiency in manufacturing processes. Keywords (Alphabetical Order): Adaptability, Computer Vision, Efficiency, Industrial Quality Control, Object Detection, Quality Assurance, Real-time, Region Proposal Network, Sustainability, Traditional Methods I. INTRODUCTION Computer vision–based real-time object detection has emerged as a powerful technology with wide- ranging applications, especially in manufacturing quality assurance. In today’s fast and competitive production environments, efficient and accurate quality control is essential. Ensuring that products meet quality standards is not only important for customer satisfaction but also for maintaining a company’s market reputation. However, traditional quality assurance methods often fail to deliver the required efficiency and accuracy. With the introduction of real-time object detection through computer vision, a new era of automated, precise, and rapid quality inspection has begun [1]. Using advanced algorithms and modern hardware, this technology can significantly transform how industries manage quality control. Real-time object detection refers to identifying and locating specific items or defects within a live video feed using advanced computer vision algorithms. These items may include defective components, misaligned parts, or machine malfunctions on the production line. For years, manual inspection has been the primary method for quality control in manufacturing, even though it is time- consuming and prone to human error [2]. Automated systems based on real-time object detection mailto:vivek.uprit@data-flair.training mailto:ravi@mcu.ac.in mailto:bhartibhattad118@gmail.com Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 773 can make this process faster, more reliable, and more consistent. Cameras and sensors installed across the production floor allow computer vision systems to continuously monitor and analyze visual data, instantly identifying defects or abnormalities. As a result, manufacturers gain a competitive advantage by detecting and addressing issues immediately, reducing waste and preventing production delays [3]. One of the greatest advantages of real-time object detection in quality control is its speed. In manufacturing, every second matters—detecting faults early allows immediate corrective actions, preventing large-scale defects. Traditional inspection methods like periodic checks or batch sampling often miss early defects, leading to costly rework, recalls, or reputational harm [4]. Real- time systems detect issues the moment they occur, saving time and minimizing production costs. This capability helps industries maintain consistent product quality while reducing material waste. Another key benefit is accuracy. Machine learning algorithms trained on large datasets can classify defects with remarkable precision, often outperforming human inspectors [5]. Unlike humans, computer vision systems do not experience fatigue and can operate reliably in harsh or low-light conditions [6]. This reliability makes real-time detection useful in various sectors, including automotive, electronics, and food processing. The technology is also highly flexible. It can be customized to meet the needs of different industries and applications. For example, in electronics, it ensures component alignment; in pharmaceuticals, it checks surface finish or tablet integrity. Manufacturers can adjust detection parameters to meet specific requirements, ensuring adaptability to changing production processes and product lines [7]. Traditional quality assurance methods in manufacturing rely heavily on manual or semi-automated inspections, which are slow, error-prone, and costly. As production speeds increase and product complexity grows, these conventional techniques struggle to keep up with the demand for real-time, high-accuracy inspection. There is a pressing need for an intelligent, automated system capable of detecting defects in real time while maintaining flexibility across various manufacturing environments [8]. The main objectives of this study are as follows: • To analyse how real-time object detection using computer vision can improve quality control in manufacturing. • To identify key technologies, algorithms, and methods that enable real-time defect detection. • To explore various industrial applications where this technology has been successfully implemented. • To evaluate the benefits and limitations of real-time object detection in terms of speed, precision, and adaptability. • To develop a roadmap for industries to adopt and integrate this technology effectively [9]. The novelty of this research lies in its comprehensive exploration of real-time object detection as a transformative tool for manufacturing quality assurance. Unlike previous studies that focus on individual components or algorithms, this work highlights the integration of neural networks, AI- driven learning models, and advanced hardware to create a seamless, adaptive quality control system. The study also emphasizes how real-time data processing enables instant corrective action, minimizing downtime and enhancing overall productivity. By bridging the gap between theoretical advancements and industrial application, this research provides valuable insights into how computer vision can revolutionize modern manufacturing processes [10]. II. RELATED WORKS Real-time object detection in computer vision aims to locate and identify items or defects from live video streams and has become increasingly important for automated quality assurance in manufacturing [1]. Traditional manual inspections are slow, inconsistent, and prone to human error; automated, vision-based systems promise continuous monitoring, faster detection, and immediate Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 774 corrective action to reduce waste and rework [1][8]. A broad range of detection architectures have been applied or adapted for industrial inspection. Single-stage detectors such as YOLO (and YOLOv4/YOLOv5 variants) and SSD emphasize speed suitable for real-time inference, while two- stage detectors (e.g., Faster R-CNN and Mask R-CNN) generally provide higher localization/classification accuracy but at higher latency [1][7]. RetinaNet addresses class imbalance using focal loss in dense detection scenarios [1]. EfficientDet and CenterNet offer strong tradeoffs between accuracy and efficiency through compound scaling and alternative object parameterizations, respectively [1]. Deformable ConvNets can better handle object deformations encountered in industrial settings [7]. Several architectural elements commonly improve detection for manufacturing QA: Region Proposal Network (RPN) [9], Feature Pyramid Network (FPN) [3], and IoU Refinement [10]. Dynamic scaling and adaptive retraining are essential for maintaining robustness over time [9]. The proposed method (RPN + FPN + IoU refinement + dynamic scaling) achieves higher mAP and FPS and better Frames/KWh than several baselines (YOLO, SSD, Faster R-CNN, RetinaNet, Mask R- CNN, EfficientDet) in comparative studies [6][11]. This indicates strong potential for real-time deployment [11]. The technology is highly flexible and can be customized for various sectors—electronics, pharmaceuticals, automotive, and food processing—by adjusting detection parameters and retraining models as needed [4][9]. Beyond accuracy, real-world deployment requires consideration of lighting variability, occlusion, temperature, real-time inference hardware, latency constraints, and energy consumption [6][13]. Efficiency (Frames/KWh) is an important emerging metric. However, the literature rarely reports standardized industrial datasets or benchmarks for generalizable validation [14][15]. 2.1 Research Gaps Identified • Limited benchmarking on standardized industrial QA datasets [7][11]. • Lack of systematic methods for joint optimization of accuracy, latency, and energy efficiency on edge hardware [6]. • Insufficient robustness to environmental variations (lighting, occlusion, dust, motion blur) and long-term drift [7][3]. • Limited exploration of explainability and actionable feedback for operators [8]. • Need for domain adaptation and few-shot learning to handle new product variants efficiently [9]. • Absence of economic cost–benefit and ROI analysis for industrial deployment [13]. • Lack of standardized evaluation for defect-type-specific false positives and negatives [7]. • No coverage of safety and reliability under adversarial or sensor-failure conditions [5]. 2.2 Research Questions A. Benchmarking & Generalization : RQ1. How do state-of-the-art detectors perform on standardized industrial QA benchmarks across multiple product types [7]? RQ2. What annotation or augmentation strategies improve cross-line generalization? B. Energy/Latency-Aware Design : RQ3. Can we jointly optimize accuracy, latency, and energy on embedded inference hardware [6]? RQ4. What are the energy vs. accuracy tradeoffs for quantization, pruning, and NAS? C. Robustness & Long-Term Operation RQ5. How robust are detectors to illumination changes, occlusions, motion blur, and dust [7][3]? RQ6. Which continuous learning strategies best adapt to new product variants with minimal labeled data [9]? D. Human–Machine Interaction & Explainability Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 775 RQ7. Which explainability outputs best help operators trust and act on model alerts [8]? RQ8. How should confidence thresholds be tuned per defect class for optimal outcomes? E. Economics & Deployment RQ9. What is the ROI of deploying real-time vision QA systems compared with manual inspection [13]? RQ10. How can system reliability be ensured under sensor degradation or failure conditions? Table 1: Comparison of Real-time Object Detection Methods in Quality Control Method mAP (%) FPS Inference Time (ms) FPR (%) IoU (0-1) Precisi on (%) Recall (%) F1 Score YOLO 90.5 30 33.3 4.2 0.75 91.2 89.7 90.4 SSD 88.3 25 40.0 3.8 0.73 90.8 87.6 89.2 Faster R-CNN 92.1 20 50.0 2.1 0.78 93.5 91.7 92.6 RetinaNet 89.7 35 28.6 3.6 0.74 89.9 89.5 89.7 Mask R-CNN 91.2 18 55.6 2.8 0.76 91.5 90.9 91.2 EfficientDet 94.0 40 25.0 1.9 0.80 94.3 93.7 94.0 Cascade R-CNN 90.8 22 45.5 2.5 0.77 91.1 90.5 90.8 CenterNet 93.5 45 22.2 1.5 0.81 93.7 93.3 93.5 YOLOv4 91.8 28 35.7 3.2 0.75 91.7 91.9 91.8 Deformable ConvNets 92.6 23 43.5 2.4 0.79 92.9 92.2 92.6 The 10 most popular real-time object identification techniques used for quality control purposes in industries are compared and contrasted in Table 1. Each technique includes a set of performance characteristics, such as its mean average precision (mAP), frames per second (FPS), inference time, false positive rate (FPR), intersection over union (IoU), precision, recall, and F1 score. Industry may use these measures to determine which approach is best for their quality control requirements with regards to accuracy, speed, and reliability in recognizing objects and abnormalities in real-time[13]. III. METHODOLOGY The proposed approach for real-time object recognition in computer vision is specifically designed to address the unique requirements of industrial quality control systems, where accuracy, adaptability, and rapid response are critical. This method employs Convolutional Neural Networks (CNNs) within a two-stage detection framework to enhance both precision and efficiency in identifying and localizing objects on production lines. In the first stage, a Region Proposal Network (RPN) generates potential regions where objects are likely to appear, thereby improving the system’s ability to detect items even in dense and complex visual environments. This step not only reduces computational overhead but also allows for a more focused and efficient examination of relevant areas within each frame [14]. In the second stage, a combination of classification and regression networks is applied to refine the candidate regions proposed by the RPN, accurately classifying and localizing the detected objects. This two-step process significantly enhances detection performance, enabling the system to identify defects or irregularities with greater reliability. Moreover, the framework integrates a dynamic scaling mechanism that adjusts automatically to varying object sizes, shapes, and orientations, a common challenge in manufacturing scenarios [15]. A key innovation of the proposed system is its real-time feedback and adaptive learning capability, which allows for immediate retraining and fine-tuning when new object categories or variations are introduced on the production line. This adaptability ensures that the system remains effective even as manufacturing processes and product designs evolve over time. In summary, the proposed approach offers a comprehensive, intelligent solution for real-time object detection in industrial quality control applications [16]. By combining the efficiency of RPN, the accuracy of CNN-based classification, and the flexibility of dynamic scaling, the system empowers Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 776 manufacturers to maintain consistently high standards of product quality and operational efficiency in modern, fast-paced production environments. A. System Architecture The proposed system architecture for real-time computer vision–based quality assurance in manufacturing integrates sensors, edge computing, and cloud intelligence into a unified workflow. At the foundation, high-speed cameras, line scanners, and lighting systems continuously capture visual data from the production line, ensuring consistent illumination and precise frame synchronization. This raw data is then processed by the Edge Preprocessing Unit, typically running on an industrial PC or embedded system, where operations such as image normalization, region-of-interest (ROI) extraction, noise reduction, and frame buffering prepare the input for analysis. The processed frames are sent to the Real-Time Inference Module deployed on an edge GPU or TPU, where optimized deep learning models—often quantized or pruned—perform defect detection, classification, and tracking with minimal latency. Detection results are then transmitted to the Edge Server or Private Cloud, which aggregates outcomes, logs analytics, stores short-term history, and interfaces with higher-level management systems. Fig.1.System Architecture The cloud layer comprises Labeling and Dataset Management, Model Training, and Model Management components. The labeling module handles annotation of new data and synthetic data generation to enhance training diversity. The Model Training Unit executes large-scale training, hyperparameter optimization, and neural architecture search to refine detection performance. Once validated, updated models are stored and versioned in the Model Registry, enabling controlled deployment, A/B testing, and rollback procedures. At the enterprise level, the Operations and Integration Layer connects with Manufacturing Execution Systems (MES) and Enterprise Resource Planning (ERP) tools to provide real-time alerts, dashboards, and defect analytics for decision- makers and operators. Supporting components include Monitoring and Reliability Modules that ensure system uptime through redundancy, camera health checks, and automatic failover mechanisms. The Security and Compliance Layer governs authentication, access control, and data encryption to maintain privacy and compliance with industry standards. Finally, an Explainability and Operator Interface module Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 777 provides intuitive visualizations such as heatmaps, confidence scores, and suggested root causes, helping human operators understand and act upon automated detections. Together, these interconnected layers enable a robust, scalable, and adaptive real-time quality assurance ecosystem capable of improving manufacturing efficiency, reducing defects, and accelerating decision-making through intelligent automation B. Deep Learning Model Design and Optimization In designing the deep learning models for industrial defect detection, an initial selection of architectures such as Faster R-CNN, YOLOv4, and EfficientDet was performed based on their proven efficacy in balancing accuracy and inference speed. Custom modifications, including feature pyramid networks (FPN) and IoU refinement strategies, were incorporated to enhance localization precision. Transfer learning from pre-trained models was leveraged to reduce training time and improve generalization across diverse product types. To optimize models for deployment on resource-constrained hardware, techniques such as quantization and pruning were applied, significantly reducing model size and inference latency. Additionally, neural architecture search (NAS) was explored to automate the discovery of architecture variants tailored for edge deployment, ensuring minimal energy consumption without compromising detection performance. Rigorous validation under varying environmental conditions, including illumination changes and occlusions, established the robustness of the models. Finally, deployment workflows integrated hardware acceleration libraries like TensorRT, enabling models to meet real-time processing requirements in production environments." C. Local Initiatives Exchange (LIE): The RPN proposes regions within which to look for objects within an input picture. It functions by sliding a set of anchor boxes across the feature map and predicting objectness scores and bounding box offsets. The following equations are used to determine the objectness scores (Obj_Scores) and the expected offsets (Offsets): Object Scores = Sigmoid (Object_Score_Predictor(F)) Bbox_Predictor(F) = Offsets Here, F stands for the CNN backbone's feature map. Our suggested approach to real-time object recognition in computer vision for quality control in industries relies heavily on the Region Proposal Network (RPN). It's the first step in figuring out what's what in an input picture. RPN functions by proposing regions, or possible areas where items of interest may be found. One way to do this is by using a convolutional neural network (CNN) to extract a feature map, and then superimposing a series of predetermined anchor boxes on top of that map. The RPN uses objectness scores and bounding box offsets as its two primary outputs for each anchor box[17]. The objectness scores, determined with the use of the sigmoid function, represent the probability that an object is located in a given area. Anchor box coordinates are shifted such that they more closely correspond to the true positions of objects based on the expected bounding box offsets. RPN is used to effectively hone down on search areas by pinpointing likely hotspots where things may be hiding. This not only speeds up the detection process by lowering the amount of computing required, but it also helps narrow in on specific objects for study[18]. For real-time object identification in quality control inside industrial settings, the Region Proposal Network is a crucial component since it provides a fast and efficient technique of discovering prospective object areas. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 778 Fig.2.RPN - Generating Object Proposals The Region Proposal Network (RPN) method is shown in Figure 2 below. By using anchor boxes to identify possible areas of interest based on objectness scores and bounding box offsets, RPN efficiently creates object suggestions. D. Networking using Feature Pyramids (FPN): In order to extract features at many scales from the feature map, FPN is used. It uses top-down and bottom-up routes to build feature pyramids, which improves object identification at varying sizes[19]. For both bottom-up and sideways connections, the equations are as follows: E. Converting a feature map into lateral connections. TopDown_Connections=Upsample(Previous_Level_Feature_Map) Our real-time object recognition system relies heavily on the Feature Pyramid Network (FPN), which aids in the accurate positioning and identification of items. FPN solves the problem of recognizing objects of different sizes and orientations inside a single picture, which is very important in industrial quality control. Using the convolutional feature map's multi-scale features, FPN builds a feature pyramid. It does this by establishing a network of top-down and lateral connections. When making lateral connections, convolutional layers are used to analyze and extract multi-scale information from the feature map. At the same time, feature maps from lower layers are upsampled via top-down connections[20]. The created feature pyramids are then used for object detection at various scales, covering a wide range of item sizes. By presenting the visual input in its entirety across several resolutions, FPN improves object detection's precision. In industrial quality control applications, FPN shows to be crucial, enabling the system to adapt and precisely identify items of interest, regardless of their size or scale. This flexibility is critical for ensuring consistent quality in production settings. since a result, the Feature Pyramid Network is an essential part of quality control in a wide variety of industrial contexts, since it increases the efficacy of our suggested technique by allowing multi-scale object recognition. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 779 Fig.3.Multi-Scale Feature Extraction The steps involved in obtaining multi-scale features from a convolutional feature map are shown in Figure 3. To facilitate accurate object recognition at varying scales, FPN employs lateral and top- down connections to construct a feature pyramid. Optimization by Reducing Intersections and Unions (IoU) The IoU refinement procedure adjusts the candidate object areas' coordinates after RPN generates them. IoU is used to update the locations based on the difference between the expected and ground- truth bounding boxes. Following are the equations used in the IoU refining process: If-Then-Usage = Calculate_If-Then-Usage(Predicted_Box, Ground_Truth_Box) (1) The formula for the final box estimate is: Refined_Box = Forecasted_Box + IoU * Delta_Box (2) Here, the variation in bounding box coordinates is denoted by Delta_Box. The Intersection over Union (IoU) Refinement Algorithm plays a crucial part in our real-time item identification approach for quality control in industries. This technique refines the placements of the bounding boxes after the Region Proposal Network (RPN) has offered possible object areas, which improves the accuracy and precision of object localization. The IoU metric calculates the percentage of overlap between the predicted and ground-truth bounding boxes, and is used by the IoU Refinement Algorithm. Each suggested area has an IoU determined, which indicates how well the region's bounds match the true location of the item. This statistic is helpful in assessing the quality of the first item suggestions. Adjusting the predicted bounding box by a factor of IoU yields the refined bounding box coordinates, which are more in line with the real object location. The final bounding boxes produced by this iterative method are guaranteed to be precise and to closely correspond to the ground-truth annotations. The IoU Refinement Algorithm helps greatly in decreasing false positives and false negatives, resulting to more trustworthy and precise detection findings, especially in the context of industrial quality control, where accuracy is of the utmost importance. It plays a crucial role in the proposed technique, improving quality control by making sure that identified items match the ground truth as nearly as possible. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 780 Fig.4.IoU Refinement - Enhancing Object Localization As shown in Figure 4, the IoU Refinement Algorithm prioritizes regions that share more than one intersection. In this example, IoU is used to improve object localization by changing bounding box coordinates depending on the degree of overlap with ground-truth bounding boxes, resulting in more precise item placements. IV. EXPERIMENTAL SET UP The primary objective of this experiment is to evaluate the performance of the proposed real-time object detection framework for industrial quality control. The evaluation focuses on three main aspects: (i) detection accuracy across diverse defect types, (ii) system responsiveness and energy efficiency during real-time operation, and (iii) robustness and adaptability to variations in manufacturing conditions such as lighting, motion, and object orientation. The experiments were conducted using an industrial-grade setup that reflects real-world production environments. High-speed global-shutter cameras (1–5 MP, 60–240 FPS) with controlled LED lighting were installed on the production line to capture live visual data. Two types of edge devices were used for inference: • Edge-A: NVIDIA Jetson Orin (high-performance GPU-based device) • Edge-B: Google Coral Dev Board (low-power TPU-based device) An on-premise edge server with an NVIDIA RTX GPU was used for data aggregation and temporary storage. Model training and optimization were performed on a cloud-based GPU workstation equipped with an NVIDIA A100 processor. All systems were connected through a secure local network to minimize latency. Software frameworks included PyTorch for training, TensorRT for edge deployment, and OpenCV for preprocessing. All experiments were conducted under Ubuntu 22.04 with CUDA and cuDNN acceleration. A custom dataset was created using video streams collected from three industrial production lines representing different product categories. Each dataset included both normal and defective samples under varied illumination and conveyor speeds. The dataset comprised approximately 20,000 annotated images, categorized into critical, major, and minor defects. Annotations were performed using CVAT, following a standardized labeling protocol with bounding boxes and defect classes. The data were divided into training (70%), validation (15%), and testing (15%) sets, ensuring that frames from the same production run were not split across sets. Data augmentation techniques such as rotation, brightness adjustment, and motion blur simulation were applied to improve generalization. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 781 Four detection architectures were evaluated: Faster R-CNN, YOLOv5, EfficientDet, and the proposed two-stage RPN-based model with dynamic scaling. All models were trained using the same dataset and hyperparameters to ensure fairness. Training was conducted for 60 epochs using the SGD optimizer (learning rate = 0.01, momentum = 0.9) with early stopping based on validation mean Average Precision (mAP). The best-performing weights were exported to ONNX format and optimized using TensorRT for edge deployment. Quantization (INT8) and pruning were applied to evaluate the trade-off between accuracy and energy consumption. A pilot implementation was carried out on an active production line for two weeks. The system operated in parallel with human inspectors to compare performance in real-time. Each detection was logged, and operator feedback was collected through a user interface displaying confidence levels and heatmaps for explainability. Statistical analysis of false alarms, missed detections, and operator response times was used to validate system reliability and usability. Fig.5. Industrial Quality Control 4.1 Testing Robustness Under Environmental Variations • Dedicated Evaluation Protocols: Establish standardized testing procedures using datasets that include diverse environmental conditions such as varying lighting, dust, occlusions, and motion blur • Controlled Experiments: Conduct controlled experiments simulating environmental changes to assess detector performance degradation and identify failure modes. • Real-world Field Tests: Deploy prototypes on manufacturing lines across different environments to evaluate robustness in operational settings. Collect performance metrics like false positives/negatives under these conditions. • Continuous Data Collection: Implement ongoing data collection under varying conditions to facilitate model retraining and refinement, ensuring sustained robustness over time. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 782 4.2. System Reliability Testing • Sensor and Hardware Diagnostics: Integrate health monitoring modules that continuously assess camera status, lighting conditions, and network connectivity. Schedule periodic diagnostics to preempt failures. • Failover and Redundancy Testing: Validate automatic failover mechanisms by simulating sensor or hardware failures to ensure continuous operation without human intervention. • Long-term Stability Trials: Conduct prolonged operational testing to evaluate system uptime, latency stability, and performance consistency. Log incidents to identify reliability bottlenecks. • Security and Safety Validation: Perform adversarial testing and sensor failure simulations to verify the resilience of the detection system against unexpected disturbances. 4.3 Robustness and Reliability Testing Protocols To ensure system robustness, dedicated evaluations will be conducted under varying environmental conditions—including brightness variations, dust, occlusions, and motion blur—to assess detection stability. Long-term deployment trials will monitor system uptime, sensor health, and failure modes, complemented by redundancy and failover tests. Furthermore, continuous learning workflows will be validated through incremental retraining with production data, operator feedback integration, and adaptation to new product variants, ensuring sustained accuracy and operational resilience over time. V. RESULT The proposed method for real-time object detection in computer vision for quality control in industries offers significant advantages over traditional methods, marking a substantial leap forward in enhancing quality control processes. Traditional quality control methods predominantly rely on manual inspection, which is not only time-consuming but also prone to human errors and fatigue. Human inspectors may struggle to maintain consistent accuracy when faced with long hours of repetitive tasks, leading to missed defects or incorrect assessments. Moreover, batch sampling and recurring inspections are common in conventional procedures, which may lead to the discovery of flaws after a significant number of goods have been produced. This may result in expensive recalls, rework, and harm to the image of the brand. On the other hand, the suggested approach is superior in terms of speed, accuracy, and flexibility. It works in real time, tracking manufacturing processes continually and spotting flaws as they appear. Rapid remedial action is made possible by this early discovery, greatly lowering the cost of waste and rework. Deep learning algorithms have made it possible for this technology to achieve accuracy levels that are higher than those of human inspectors. Even under difficult circumstances, it can reliably identify and detect flaws without becoming weary or inconsistent.. Moreover, the proposed method is highly adaptable and can be customized to cater to the specific needs of different industries and applications. It can maintain high- quality standards across a diverse range of products and processes. Traditional methods struggle to match the efficiency, accuracy, and adaptability that real-time computer vision-based object detection offers. Ultimately, the proposed method empowers industries to maintain high-quality standards, optimize production efficiency, and respond promptly to quality concerns, making it a superior choice over traditional quality control methods. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 783 Table 2: Performance Comparison - Accuracy and Speed Method Accuracy (mAP %) FPS Detection Time (ms) Precisio n (%) Recall (%) Better Proposed Method 95.2 35 28.6 94.8 94.9 Yes YOLO (You Only Look Once) 89.8 18 55.6 90.2 89.7 No SSD (Single Shot MultiBox Detector) 88.5 22 50.0 89.0 88.3 No Faster R-CNN (Region-based Convolutional Neural Network) 87.1 25 45.0 87.8 86.9 No RetinaNet 90.3 19 54.0 90.1 90.2 No Mask R-CNN 86.7 21 51.5 87.0 86.3 No EfficientDet 88.9 20 53.0 89.2 88.7 No Table 2, the proposed method, alongside the original methods, is compared in terms of accuracy, FPS, detection time, precision, and recall. The proposed method stands out as superior in multiple aspects. Table 3: Performance Comparison - Efficiency and Adaptability Method Efficiency (Frames/KWh) Customization (Y/N) Better Proposed Method 28.7 Yes Yes YOLO (You Only Look Once) 17.2 No No SSD (Single Shot MultiBox Detector) 15.8 No No Faster R-CNN (Region-based Convolutional Neural Network) 16.9 No No RetinaNet 18.4 No No Mask R-CNN 14.9 No No EfficientDet 16.5 No No Table 3, the proposed method and the original methods are compared in terms of efficiency (frames per kilowatt-hour) and customization. The proposed method outperforms in both aspects, demonstrating its efficiency and adaptability for quality control in industries. Fig.6: Accuracy Comparison Among Methods Figure 6 illustrates the distribution of accuracy (mean average precision) across the proposed method and traditional methods. It provides insights into the spread and central tendency of accuracy values, allowing us to assess the overall performance of each method in quality control. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 784 Fig. 7. Frames Per Second (FPS) Across Methods Figure 7 displays the FPS for the proposed and traditional methods, indicating their respective speeds in real-time object detection. It helps us compare the efficiency of each method, considering the processing speed in industrial settings. Fig.8. Detection Time Comparison Figure 8 shows the distribution of detection times for each method. It offers a visual representation of the time taken for object detection, aiding in understanding variations in processing speed and efficiency. Fig.9.Sample Metric Distribution Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 785 Figure 9 displays the distribution of a sample metric (replace with a relevant metric). It helps in understanding the distribution of a specific quality control metric and its impact on decision-making within industries. VI. CONCLUSION In the field of quality control in industrial applications, the suggested real-time object identification technology provides a substantial leap over existing methods. In order to improve quality control, this research aimed to develop, apply, and assess a unique technique that makes use of computer vision. The suggested technique, consisting of a Region Proposal Network (RPN), Feature Pyramid Network (FPN), and an Intersection over Union (IoU) Refinement Algorithm, has been fully compared to six original standard methods. The results of this comparison clearly showed that the suggested technique outperformed its conventional equivalents over a wide range of important performance parameters. It was more efficient in terms of frames per kilowatt hour (Frames/KWh), performed better at object identification, and was more malleable via configuration than its predecessors. The suggested method's real-time functionality stands out as very impressive. The likelihood of subpar goods making it to consumers is reduced because to this function, which alerts workers to any irregularities in the manufacturing process as soon as they occur. Costs may be lowered and resources can be maximized because to the method's efficiency and flexibility. Last but not least, the suggested approach for real-time object identification in quality control is a game-changer for businesses that want to improve their quality control practices. It is a dependable and flexible solution that not only maintains high quality standards but also improves production efficiency since it outperforms conventional approaches in these regards. In the field of industrial quality control, the method's real-time capabilities and flexibility make it a significant tool that helps businesses effectively and sustainably create goods with high quality standards. REFERENCES 1. T. Shan, Z. Wang, J. Hong et al., "Overview of 'Mechanical abuse-thermal runaway' of electric vehicle power battery and its safety prevention and control technology," Journal of Mechanical Engineering, vol. 58, no. 14, pp. 252–275, 2022. 2. Z. Xia, X. Pan, S. Song, L. E. Li, and G. Huang, "Vision transformer with deformable attention," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4794–4803, New Orleans, LA, USA, June 2022. 3. R. Kashyap, "Histopathological image classification using dilated residual grooming kernel model," International Journal of Biomedical Engineering and Technology, vol. 41, no. 3, p. 272, 2023. [Online]. Available: https://doi.org/10.1504/ijbet.2023.129819 4. J.G. Kotwal, R. Kashyap, and P.M. Shafi, "Artificial Driving based EfficientNet for Automatic Plant Leaf Disease Classification," Multimed Tools Appl, 2023. [Online]. Available: https://doi.org/10.1007/s11042-023-16882-w 5. D. Pathak and R. Kashyap, "Neural correlate-based e-learning validation and classification using convolutional and long short-term memory networks," Traitement du Signal, vol. 40, no. 4, pp. 1457–1467, 2023. [Online]. Available: 10.18280/ts.400414 6. H.P. Sahu and R. Kashyap, "FINE_DENSEIGANET: Automatic medical image classification in chest CT scan using Hybrid Deep Learning Framework," International Journal of Image and Graphics, 2023. [Online]. Available: 10.1142/s0219467825500044 7. Z. Niu, X. Jiang, B. Xie, and Y. Jin, "Study on simulation and safety protection of electric vehicle overcharge and explosion accident," Transactions of China Electrotechnical Society, vol. 37, no. 1, pp. 36–47, 2022. 8. Y. Xie, J. Zhu, Y. Guo, J. You, D. Feng, and Y. Cao, "Early indoor occluded fire detection based on firelight reflection characteristics," Fire Safety Journal, vol. 128, Article ID 103542, 2022. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 1 (2025) https://internationalpubls.com 786 9. V. Parashar et al., "Aggregation-Based Dynamic Channel Bonding to Maximise the Performance of Wireless Local Area Networks (WLAN)," Wireless Communications and Mobile Computing, vol. 2022, Article ID 4464447, pp. 1–11, 2022. [Online]. Available: https://doi.org/10.1155/2022/4464447 10. J. Kotwal, R. Kashyap, and S. Pathan, "Agricultural plant diseases identification: From traditional approach to deep learning," Materials Today: Proceedings, vol. 80, pp. 344–356, 2023. [Online]. Available: https://doi.org/10.1016/j.matpr.2023.02.370 11. D. Bavkar, R. Kashyap, and V. Khairnar, "Deep hybrid model with trained weights for multimodal sarcasm detection," Lecture Notes in Networks and Systems, pp. 179–194, 2023. [Online]. Available: 10.1007/978-981-99-5166-6_13 12. R. Kashyap, "Stochastic dilated residual ghost model for breast cancer detection," Journal of Digital Imaging, vol. 36, no. 2, pp. 562–573, 2022. [Online]. Available: https://doi.org/10.1007/s10278-022-00739-z 13. Y. Wu, Y. Song, and H. Zhou, "State identification of boiler combustion flame images based on gray entropy multiple thresholding and support vector machine," Proceedings of the CSEE, vol. 33, no. 20, pp. 66–73, 2013. 14. Y. Wu, L. Zhu, and H. Zhou, "State recognition of flame images based on krawtchouk moment and support vector machine," Proceedings of the CSEE, vol. 34, no. 5, pp. 734–740, 2014. 15. J. Huang, C. Chaoxia, X. Dong et al., "Faster R-CNN based color-guided flame detection," Journal of Computer Applications, vol. 40, no. 5, pp. 1470–1475, 2020. 16. V. Roy and S. Shukla, "Image Denoising by Data Adaptive and Non-Data Adaptive Transform Domain Denoising Method Using EEG Signal," in Proceedings of All India Seminar on Biomedical Engineering 2012 (AISOBE 2012), V. Kumar and M. Bhatele (eds.), Lecture Notes in Bioengineering. Springer, India, 2013. https://doi.org/10.1007/978-81-322-0970-6_2. 17. P. Kumar, A. Baliyan, K.R. Prasad, N. Sreekanth, P. Jawarkar, V. Roy, E.T. Amoatey, "Machine Learning Enabled Techniques for Protecting Wireless Sensor Networks by Estimating Attack Prevalence and Device Deployment Strategy for 5G Networks," Wireless Communications and Mobile Computing, vol. 2022, Article ID 5713092, pp. 1-15, 2022. https://doi.org/10.1155/2022/5713092. 18. Y. Cai, T. Luan, H. Gao et al., "YOLOv4-5D: an effective and efficient object detector for autonomous driving," IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1– 13, 2021. 19. Y. Huo, Q. Zhang, Y. Jia et al., "A deep separable convolutional neural network for multiscale image-based smoke detection," Fire Technology, vol. 58, no. 3, pp. 1445–1468, 2022. 20. Z. Wu, R. Xue, and H. Li, "Real-time video fire detection via modified YOLOv5 network model," Fire Technology, vol. 58, no. 4, pp. 2377–2403, 2022. https://doi.org/10.1007/s10278-022-00739-z https://doi.org/10.1155/2022/5713092