<p>The massive scale of deep learning models limits their application on edge devices. Therefore, reducing the number and complexity of model parameters has become a research hotspot. The existing model pruning methods do not consider the interlayer coupling relationship between channels, resulting in the loss of some correlation information and low detection accuracy. Therefore, we propose an interlayer coupled pruning method based on elastic scaling score factors. Firstly, considering that the current layer filter may have an impact on subsequent channels, an intra layer coupled channel pruning method is proposed by adding inter layer coupling relationships to determine whether the channel is redundant based on channel pruning. Secondly, in the process of sparsification, to more accurately evaluate the importance of channels, L1-norm and L2-norm were combined for sparse regularization, and an elastic scaling scoring factor <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\gamma \)</EquationSource> </InlineEquation> is proposed. Finally, considering that small objects in actual traffic scene are highly susceptible to interference from rich background information, a combination of WIOU+NWD regression loss function is proposed to improve the stability and detection accuracy of the model training process. Through sufficient experimental verification, on the KITTI public transportation dataset, with almost unchanged <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\varvec{mAP}\)</EquationSource> </InlineEquation> compared to the benchmark network YOLOv10n, the <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\varvec{Params}\)</EquationSource> </InlineEquation> decreased by 29.1%, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\varvec{FLOPs}\)</EquationSource> </InlineEquation> decreased by 5.6%, and detection speed increased by 1190.5<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\varvec{FPS}\)</EquationSource> </InlineEquation>. On the Cityscapes dataset, compared with the benchmark model YOLOv10x, without a decrease in <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\varvec{mAP}\)</EquationSource> </InlineEquation> value, the <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(\varvec{Params}\)</EquationSource> </InlineEquation> decreases by 36.6% and the speed increases by 67.9%.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

ICCPNet: inter-layer coupling channel pruning network based on elastic scaling scoring for complex traffic scene

  • Cuijin Li,
  • Jianyu Liu,
  • Zhong Qu

摘要

The massive scale of deep learning models limits their application on edge devices. Therefore, reducing the number and complexity of model parameters has become a research hotspot. The existing model pruning methods do not consider the interlayer coupling relationship between channels, resulting in the loss of some correlation information and low detection accuracy. Therefore, we propose an interlayer coupled pruning method based on elastic scaling score factors. Firstly, considering that the current layer filter may have an impact on subsequent channels, an intra layer coupled channel pruning method is proposed by adding inter layer coupling relationships to determine whether the channel is redundant based on channel pruning. Secondly, in the process of sparsification, to more accurately evaluate the importance of channels, L1-norm and L2-norm were combined for sparse regularization, and an elastic scaling scoring factor \(\gamma \) is proposed. Finally, considering that small objects in actual traffic scene are highly susceptible to interference from rich background information, a combination of WIOU+NWD regression loss function is proposed to improve the stability and detection accuracy of the model training process. Through sufficient experimental verification, on the KITTI public transportation dataset, with almost unchanged \(\varvec{mAP}\) compared to the benchmark network YOLOv10n, the \(\varvec{Params}\) decreased by 29.1%, \(\varvec{FLOPs}\) decreased by 5.6%, and detection speed increased by 1190.5 \(\varvec{FPS}\) . On the Cityscapes dataset, compared with the benchmark model YOLOv10x, without a decrease in \(\varvec{mAP}\) value, the \(\varvec{Params}\) decreases by 36.6% and the speed increases by 67.9%.