Academic Journal of Science and Technology 
ISSN: 2771-3032 | Vol. 2, No. 3, 2022 

 

1 

Semantic	Segmentation	Based	on	Improved	Robinson	
Operator	
Xiaoshuo Jia*, Zhihui Li 

School of Computer Science, Guangdong University of Science and Technology, Dongguan 523079, Guangdong, China 

 

Abstract: Classical pixel-level semantic segmentation methods are easily affected by factors such as environment and pixel 
values, resulting in low accuracy of their methods. Therefore, in order to improve the accuracy of pixel-level semantic 
segmentation under practical conditions, this paper designs a lightweight network based on the nonlinear characteristics of 
residual structure and the characteristics of optimized edge operators. The experimental results show that the designed 
lightweight residual network can effectively improve the accuracy by about 2% compared with the traditional residual network, 
and the model size is reduced from the original 40.2MB to 18.9MB. 

Keywords: Semantic segmentation, Residual structure, Pooling layer, Lightweigth model, Robinson Operator. 
 

1. Introduction 
In image processing, pixel level semantic segmentation [1-

3] is an extremely important and complex task. CNN 
algorithm [4-7] has made excellent achievements in image 
classification, segmentation and tracking of kaggle and AI 
Challenger competitions by taking advantage of the 
characteristics of multi parameters. FCN uses deconvolution 
for up sampling to make the extracted features more detailed. 
U-net uses network symmetric structure to fuse high-
dimensional features and low-dimensional features to weight 
edge features. In cpfnet, the concept of dilated convolution is 
proposed, which can expand the field of view of convolution 
layer to extract more feature information. Then, the context 
based feature fusion is realized in combination with the 
inception module, which has achieved excellent results in 
medical data sets. STDC is based on FPN for multi-scale 
fusion, so its performance is better than cpfnet algorithm. 
Although the algorithm based on CNN has high accuracy, the 
extracted features usually lose a lot of spatial information due 
to the pool layer. Finally, problems such as redundancy of 
semantic segmentation network structure, large amount of 
calculation and segmentation errors appear. 

Here, we analyze the statistical properties and distribution 
properties of the data on the data distribution of the image 
edge features, and design a correlation pooling layer rel layer. 
Then the NR layer is designed on the basis of Robinson 
operator to extract the edge contour features. Finally, renrnet 
is designed by combining NR layer and rel layer in residual 
structure. Renrnet and SOTA algorithms were compared 
under the SBD dataset. The experimental results show that 
ReNRnet has certain advantages in accuracy and speed. 

2. Method 

2.1. Optimized Edge Operator NR 
On the basis of the Robinson operator, an optimized edge 

operator NR is designed, at the orientation of the polar 
coordinate system of the image, and 8 polar coordinate 

convolution kernels are given. The picture pic and the 8 
azimuth convolution kernels R1...R8 do the convolution 
operation and then accumulate the result cov, such as formula 
(1): 

8

1

c o v  =  *i
i

R p i c



 
(1) 

Obtain the minimum value min and maximum value max 
in cov, and then perform the operation, and finally obtain the 
pixel value Pm, i, j of the edge feature image P. m, i, and j 
represent the width, length, and pass numbers of the image, 
respectively. 

2.2. ReNRnet 
In feature extraction, NR can effectively extract the spatial 

position of edge features and reduce the interference of 
unnecessary features. Here, to ensure that the correlation 
between features is not lost during forward propagation, we 
propose a Rel layer, as formula (2) and (3). 

 
,

, , , ,, , ,
( * ) * ) / *m n m n m n

i j i j i j i ji j i j i j
S p p p p h w   （  (2) 

  
, , ,

,
m n m m n n

m nr S S S  (3) 

  
The feature map P obtains n special detection maps P1,..., 

PN (only five feature maps are illustrated here) under the 
sliding window movement with the step size s of 1. P2,..., PN 
all intersect with P1. P1, P2,..., P5 get the corresponding 
correlation features R1, 2,..., R1, 5 under formula 2 and formula 
3, and then regression the correlation features R1, 2,..., R1, 5 to 
the positions where P1, P2,..., P5 interact with each other, and 
finally get the correlation feature map rel1, 2,..., rel1, 5. P1 and 
the surrounding characteristic map P2,..., P5 are calculated by 
algorithm2 to obtain the correlation coefficient characteristic 
map relp1 of P1. The eigenvalue of relp1 indicates the degree 
of correlation between P1 and the surrounding feature maps 
P2,..., P5. Finally, the result of point multiplication is regressed 
to the spatial position corresponding to P1. 

 
 
 
 



 

2 

Table 1. The network structure parameters are shown in the following table 

Layer Kernel/stride 
Conv1 128*11*11/4 

NR 5*3/15 

Conv3/5/7/9/11/13 
16*1*3/2 
16*3*1/2 

Rel pool 3*3/2 
Conv2/4/6/8/10/12 8*3*3/2 

 
It can be seen from the principle that the Rel layer can not 

only effectively filter the interference of non-related features 
and improve the accurate removal rate, but also strengthen the 
features according to a certain proportion, thus helping NR 
improve the efficient positioning of edge features. Finally, 
ReNRnet achieves efficient semantic segmentation effect. 

3. Experiments 

3.1. Dataset  
The Pacal VOC (07 + 12) dataset is divided into 4 

categories and 20 subcategories. There are 33K photos in total, 
including 16K train-val data set and 16K test data set. 

SBD belongs to the enhanced dataset of VOC dataset. It 
contains 11355 marker images in the VOC. We used 
segmentation of VOC datasets as annotations to train the 
network. SBD divided the whole data set into two parts, 8498 

training images and 2857 test images.  

3.2. Comparison with SOTA  
The training platforms are RTX 3060. The optimization 

function is Adam optimizer. The loss function uses formula 4 
to calculate the error between the true value and the predicted 
value, and uses IoU to evaluate the test results. yp and yt 
represent the predicted values and actual values respectively.  

 
dice p, t  2 ∗ |𝑦 ∩ 𝑦 | 𝑦 |𝑦 |⁄  (4) 

  
Here we compare Resnet with some SOTA algorithms, 

such as Deep snake[8], Unet[9], PANet[10], FCIS[11], ESE[12-13], 
etc. The results for the SBD datasets are shown in Table 2 
below. 

 
Table 2. Comparison results of SBD datasets. 

Network DisCnet Deep UNet PANet FCIS ESE SegNet 
AUC(%) 42.6 40.7 34.4 40.1 36.4 28.2 33.6 

Fps 28.6 20.1 19.6 23.1 17.4 15.9 20.3 
 

The SBD dataset is an enhanced dataset for Pascal VOC. 
We used the model trained under Pascal VOC as the pre 
training model and tested it under the SBD dataset. The 
results are shown in Table 2, and the segmentation effect is 
shown in Figure 1. Under the SBD dataset, the model of ESE, 
UNet and SegNet are relatively small. ReNRnet adopts a 
residual structure based on correlation features, which can 

effectively reduce the weight of the model and effectively 
improve the calculation speed. It can be seen from the 
segmentation effect in Figure 1 that whether it is a complex 
background or a small target, ReNRnet can extract the related 
features through the NR and the Rel layer under the residual 
structure, thereby realizing the lightweight model and pixel 
level segmentation. 

 

ReNRnet ESE DeepImage  
Figure 1. Image represent the original image. And ReNRnet, ESE, Deep Sanke correspond to the segmentation effect image 

of these algorithms respectively. 
 



 

3 

 

4. Conclusion 
In this paper, the statistical features of discrete data and 

Robinson operator are analyzed to extract the correlation 
features in the image, and the corresponding correlation 
pooling layer and NR layer are designed. Then combined with 
residual structure, ReNRnet is designed. Then we make a 
comprehensive comparison of some SOTA algorithms under 
the SBD dataset, and prove that ReNRnet performs well in 
accuracy and speed. 

Acknowledgment 
This work is supported by Natural Science Program of 

Guangdong University of Science and Technology under the 
Grant No.GKY-2021KYQNK-2. 

References 
[1] Liu S, Qi L, Qin H, et al. Path aggregation network for instance 

segmentation[C]//Proceedings of the IEEE conference on 
computer vision and pattern recognition. 2018: 8759-8768. 

[2] Li Y, Qi H, Dai J, et al. Fully convolutional instance-aware 
semantic segmentation[C]//Proceedings of the IEEE 
conference on computer vision and pattern recognition. 2017: 
2359-2367. 

[3] Xu W, Wang H, Qi F, et al. Explicit shape encoding for real-
time instance segmentation[C]//Proceedings of the IEEE/CVF 
International Conference on Computer Vision. 2019: 5168-
5177. 

[4] Yuan Z W, Zhang J. Feature extraction and image retrieval 
based on AlexNet[C]//Eighth International Conference on 
Digital Image Processing (ICDIP 2016). SPIE, 2016, 10033: 
65-69. 

[5] Iandola F N, Han S, Moskewicz M W, et al. SqueezeNet: 
AlexNet-level accuracy with 50x fewer parameters and< 0.5 
MB model size[J]. arXiv preprint arXiv:1602.07360, 2016. 

[6] Szegedy C ,  Vanhoucke V ,  Ioffe S , et al. Rethinking the 
Inception Architecture for Computer Vision[J]. IEEE, 
2016:2818-2826. 

[7] Howard A G, Zhu M, Chen B, et al. Mobilenets: Efficient 
convolutional neural networks for mobile vision 
applications[J]. arXiv preprint arXiv:1704.04861, 2017. 

[8] Peng S, Jiang W, Pi H, et al. Deep snake for real-time instance 
segmentation[C]//Proceedings of the IEEE/CVF Conference 
on Computer Vision and Pattern Recognition. 2020: 8533-8542. 

[9] Zhao X, Vemulapalli R, Mansfield P A, et al. Contrastive 
learning for label efficient semantic 
segmentation[C]//Proceedings of the IEEE/CVF International 
Conference on Computer Vision. 2021: 10623-10633. 

[10] Armato S G ,  Roberts R Y ,  Mcnitt-Gray M F , et al. The 
Lung Image Database Consortium (LIDC) and Image Database 
Resource Initiative (IDRI): a completed reference database of 
lung nodules on CT scans.[J]. Academic Radiology, 2007, 
14( 12):1455-1463. 

[11] Jetley S, Sapienza M, Golodetz S, et al. Straight to shapes: real-
time detection of encoded shapes[C]//Proceedings of the IEEE 
conference on Computer Vision and Pattern Recognition. 2017: 
6550-6559. 

[12] Ze Yang, Yinghao Xu, Han Xue, Zheng Zhang, Raquel Urtasun,  
Liwei Wang, Stephen Lin, and Han Hu. Dense reppoints: 
Representing visual objects with dense point sets. arXiv 
preprint arXiv:1912.11473, 2019.  

[13] Xingyi Zhou, Dequan Wang, and Philipp Kr¨ahenb¨uhl. 
Objects as points. Arxiv preprint arxiv:1904.07850, 2019.