BiSeNet轻量语义分割网络优化研究
Research on BiSeNet Lightweight Semantic Segmentation Network Optimization
摘要: 语义分割是对图片上每一个像素的归类预测，使得每个语义类别对应的预测区域得以分割显现，是图像处理的重要方面。轻量级语义分割模型的研究点在于掌握性能与速度的天平，使其能够投入移动设备的应用，本文是对BiSeNet轻量语义分割模型的优化研究。首先，本文介绍了BiSeNet模型的ResNet50主体上下文分支结构，以及表层卷积辅助分支结构，还有基于通道注意力机制的ARM特征加强模块和FFM融合模块作用和原理；然后，提出模型优化改进结构，先在辅助分支表层卷积中以空洞卷积增强信息整体分析后，然后以SAM空间注意力模块增强特征质量，再利用ASPP金字塔加强辅助分支与主分支融合；最后，在VOC2012数据上，得出改进前后BiSeNet模型对比结果，在轻量性和正确性上，验证优化结构合理性。
Abstract: Semantic segmentation is an important aspect of image processing, which is to classify and predict every pixel in an image, and to segment and display the prediction region corresponding to each semantic category. The research point of light-weight semantic segmentation model is to grasp the balance of performance and speed so that it can be applied to mobile devices. Firstly, this paper introduces the RESNET50 subject context branch structure of BiSeNet model and the surface convolution auxiliary branch structure, then, the function and principle of ARM feature enhancement module and FFM fusion module based on channel attention mechanism are put forward, in the surface convolution of auxiliary branches, the whole information is enhanced by void convolution, then the feature quality is enhanced by Sam spatial attention module, and then the fusion of auxiliary branches and main branches is enhanced by ASPP pyramid, based on the VOC2012 data, the comparison results of BiSeNet model before and after improvement were obtained, and the rationality of the optimized structure was verified in the light weight and correctness.
文章引用：张梦真. BiSeNet轻量语义分割网络优化研究[J]. 计算机科学与应用, 2024, 14(4): 316-326. https://doi.org/10.12677/csa.2024.144101

参考文献

[1] Wang, Z., Tang, C., Sima, X., et al. (2021) Research on Application of Deep Learning Algorithm in Image Classification. 2021 IEEE Asia-Pacific Conference on Image Proceedings, Electronics and Computers (IPEC), Dalian, 14-16 April 2021, 1122-1125. [Google Scholar] [CrossRef] 
[2] Long, J., Shelhamer, E. and Darrell, T. (2015) Fully Convolutional Networks for Semantic Segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattren Recognition, Boston, 7-12 June 2015, 3431-3440. [Google Scholar] [CrossRef] 
[3] He, K., Zhang, X., Ren, S., et al. (2016) Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, 27-30 June 2016, 770-778. [Google Scholar] [CrossRef] 
[4] Howard, A.G., Zhu, M., Chen, B., et al. (2017) Mobilenets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv preprint arXiv: 1704.04861.
[5] Chen, L.C., Papandreou, G., Kokkinos, I., et al. (2018) DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40, 834-848. [Google Scholar] [CrossRef] 
[6] Yu, C.Q., Wang, J.B., Chao, P., et al. (2018) BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmantation. In: Ferrari, V., Hebert, M., Sminchisescu, C. and Weiss, Y., Eds., Proceedings of the 15th European Conference on Computer Vision, 334-349. [Google Scholar] [CrossRef] 
[7] Zhao, H.S., Qi, X.J., Shen, X.Y., et al. (2018) ICNet for Real-Time Semantic Segmentation on High-Resolution Images. In: Ferrari, V., Hebert, M., Sminchisescu, C. and Weiss, Y., Eds., Proceedings of the 15th European Conference on Computer Vision, 418-434. [Google Scholar] [CrossRef] 
[8] Tang, X.Y., Tu, W.X., Li, K.Q., et al. (2021) DFFNet: An IoT-Perceptive Dual Feature Fusion Network for General Real-Time Semantic Segmentation. Information Sciences, 565, 326-343. [Google Scholar] [CrossRef] 
[9] 张铮, 钱勤建, 周嘉政, 等. 基于改进Fast-SCNN的裂缝实时分割算法[J]. 应用光学, 2023, 44(3): 539-547.
[10] 方家吉, 赖一波, 唐正涛, 等. 基于DeepLabV3 的轻量级电力线语义分割方法[J]. 计算机时代, 2023(9): 19-23, 28.
[11] 肖哲璇, 陈辉, 王硕. 基于双分支多尺度特征融合的道路场景语义分割[J]. 宁夏师范学院学报, 2024, 45(1): 81-92.
[12] 谢刚, 王荃毅, 谢新林, 等. 融合多尺度深度卷积的轻量级Transformer交通场景语义分割算法[J]. 通信学报, 2023, 44(10): 212-225.