无人机语义分割双分支范围松弛匹配学习

Semantic segmentation dual⁃branch range relaxation matching learningfor unmanned aerial vehicles

先前工作在双分支网络利用蒸馏的方式使具有较高语义信息的语义分支引导空间分支学习,但是由于语义分支和空间分支之间仍然存在差异性,在部分数据集上效果不理想。为了解决该问题,文中提出范围松弛匹配学习策略,通过松弛匹配方式使学生模型不用费力匹配教师的精准输出结果,从而减小空间分支和语义分支之间的差异。相较于之前精准知识传递方式,采用范围松弛匹配方式,避免出现精准匹配对空间分支学习过于苛刻的现象,也避免了线性匹配出现过于松弛导致学习过程中梯度爆炸的现象。为了验证文中方法的有效性和泛化性,将在BiSeNetV1、BiSeNetV2、STDC三个双分支网络上进行验证,凸显该方法的有效性。通过实验结果表明文中方法较基线网络、双分支共享引导式学习、松弛匹配等方法,在UVAid和UDD两个数据集上都有提升效果,体现出范围松弛匹配的泛化性。通过对比实验表明文中改进后的方法较其他方法有一定的竞争力。

In previous research,a distillation method was utilized in a dual⁃branch network to facilitate spatial branch learning by incorporating a semantic branch with richer semantic information.However,these performance results were unsatisfactory on a part of datasets due to inherent differences between the semantic and spatial branches.In view of this,this paper introduces a learning strategy called range relaxation matching.This strategy enables the student model to closely approximate the teacher's accurate output with minimal effort,effectively minimizing the discrepancy between the semantic and spatial branches.In contrast to the previous method involving precise knowledge transfer,this paper employs a range relaxation matching technique to alleviate the rigid requirement of precise matching in spatial branch learning.This method prevents excessive relaxation caused by linear matching,which can lead to gradient explosion during the learning process.Validation experiments are conducted on three dual⁃branch networks named BiSeNetV1,BiSeNetV2 and STDC to assess the effectiveness and generalization performance of the proposed method.These experiments serve to highlight the efficacy of the proposed method.Experimental results show that the method is better than those methods of baseline networks,dual⁃branch shared⁃guidance learning and relaxation matching on the two datasets of UAVid and UDD,demonstrating the generalizability of the range relaxation matching.Comparative experiments demonstrate the competitive nature of the proposed method in comparison with the other methods.