Abstract: Abstract: In order to solve the problem that DeepLabV3+ uses holes with different dilation rates, which leads to high computational complexity and low segmentation rate for some categories, this paper proposes a deep dilation bottleneck convolution ENRF (Evolutionary Nested Receptive Fields) and adaptive class-channel attention mechanism ACCA (Adaptive Class-Channel Attention) to improve the computational complexity and segmentation rate for some categories. Firstly, by combining the two mechanisms of channel adaptive attention and class adaptive attention to construct ACCA, the feature dependencies between channels and categories are fully explored, and the expression ability of key information in the feature map is effectively enhanced. Secondly, by introducing different convolution kernel sizes and dilation rates, a deep dilation bottleneck convolution ENRF based on a nested evolution network structure of receptive field is constructed, which gradually expands the receptive field of the feature map, so as to better capture multi-scale contextual information and fine-grained edge features. Finally, the ASPP module of DeepLabV3+ is replaced by the ENRF module and ACCA is added to the fused features to improve the feature expression ability. The model can achieve continuous expansion of the receptive field in the process of feature extraction, and the feature expression is more detailed. At the same time, the overall parameter amount and computational overhead of the model are effectively reduced, making DeepLabV3+ lighter. The improved DeepLabV3+ is compared with FCN8, PSPNet, UPerNet and DeepLabV3+ in terms of FLOPs, parameters (Params), and mean intersection over Union (mIoU). The experimental results show that the improved DeepLabV3+ reduces the number of parameters, reduces FLOPs, speeds up the inference speed, and improves the segmentation performance.

Key words: Nested evolution, Lightweighting, Feature dependency, Expansion rate, DeepLabV3+

摘要： 摘 要: 针对DeepLabV3+使用不同膨胀率的空洞导致计算复杂度大和对于某些类别分割率较低的问题，提出了进化式嵌套感受野模块ENRF(Evolutionary Nested Receptive Fields)和自适应类别通道注意力机制ACCA(Adaptive Class-Channel Attention)改进方法。首先，通过结合通道自适应注意力与类别自适应注意力两种机制构建ACCA，充分挖掘通道间和类别间的特征依赖关系，有效增强了特征图中关键信息的表达能力。其次，通过引入不同卷积核大小和膨胀率构建了一种基于感受野的嵌套演化网络结构的进化式嵌套感受野模块ENRF，逐步扩大特征图的感受野，从而更好地捕捉多尺度的上下文信息及细粒度的边缘特征。最后，将DeepLabV3+的ASPP模块替换成ENRF模块并在融合后的特征中加入ACCA提高特征表达能力，模型在特征提取的过程中能够实现感受野的连续扩展，特征表达更加细致，同时有效降低了模型整体的参数量和计算开销，使DeepLabV3+更轻量化。将改进后的DeepLabV3+与FCN8、PSPNet、UPerNet和DeepLabV3+在FLOPs、参数量（Params）、均值交并比（Mean Intersection over Union，mIoU）、推理速度和内存占用五个指标上进行比较。实验结果表明，改进后的DeepLabV3+减少参数量，降低了FLOPs，加快了推理速度，提升了分割性能。

关键词: 嵌套演化, 轻量化, 特征依赖, 膨胀率, DeepLabV3+