【摘要】 针对现有语义分割算法在自动驾驶场景下精度与实时性难以兼顾的问题，提出一种结合剪枝技术与知识蒸馏技术的轻量级语义分割算法。模型结构为双路分支结构，语义分支使用ResNet50作为骨干网络提取图像语义特征，加入空洞空间金字塔模块进一步提取特征，细节分支使用3个卷积层堆叠提取图像细节信息，将两路分支信息输入多特征融合模块MFFM,融合上采样后获取分割结果，使用剪枝技术对模型冗余参数进行裁剪，结合解耦知识蒸馏技术对模型精度进行恢复提升。实验结果表明，在Cityscape数据集上，模型以10.26 M参数量，76.3%的mIoU,达到了64.16 ms的推理时间，相较于BiSeNetV2模型，在精度上高3.09%,参数量多6.9 M,推理时间慢18.74 ms。更多还原

【Abstract】 In response to the challenge of balancing accuracy and real-time performance of existing semantic segmentation algorithms in the context of autonomous driving, a lightweight semantic segmentation algorithm combining pruning and knowledge distillation techniques is proposed. The model structure consists of a dual-branch architecture. The semantic branch employs ResNet50 as the backbone network to extract semantic features from the images, incorporating Atrous Spatial Pyramid Pooling(ASPP) modules to further extract features. The detail branch uses three stacked convolutional layers to extract detailed information from the images. The information from both branches is input into a Multi-Feature Fusion Module(MFFM) for fusion and upsampling to obtain the segmentation results. Pruning techniques are applied to trim redundant parameters of the model, and knowledge distillation techniques are combined to restore and enhance the model accuracy. Experimental results demonstrate that on the Cityscape dataset, the model achieves a parameter size of 10.26 million, 76.3% mIoU, and an inference time of 64.16 milliseconds. Compared to the BiSeNetV2 model, it achieves a 3.09% higher accuracy with 6.9 million more parameters and 18.74 milliseconds slower inference time.更多还原