基于四通道不可分加性小波与DeepLabv3+结合的语义分割模型

Semantic Segmentation Model Based on Four Channel Non-Separable Additive Wavelet Combined with DeepLabv3+

摘要: 为了改善传统语义分割模型中因丢失细节，从而导致信息下降的问题，我们提出了一种改进的DeepLabv3+网络分割模型。首先将主干网络替换为MobileNetV2网络；其次通过构造四通道不可分小波低通滤波器，对源图像进行分解，提取源图像的高频子图；再次，将普通卷积更换为深度可分离卷积并且引入卷积注意力模块(CBAM)自适应细化特征，从而提高网络模型的分割效果。实验结果表明，改进后的模型在VOC数据集上均交并比(mean intersection over union, MIoU)比原始的DeepLabv3+模型提高0.94%，平均像素精度(mean pixel accuracy, MPA)比原始DeepLabv3+模型提高了1.34%，准确度比原始DeepLabv3+模型提高0.19%。在BDD100K数据集上均交并比比原始的DeepLabv3+模型提高0.53%，平均像素精度比原始DeepLabv3+模型提高了0.15%，准确率比原始DeepLabv3+模型提高0.13%。在主观和客观结果上均显示我们的模型优于原模型。

Abstract: In order to improve the loss of details in the traditional semantic segmentation model, which leads to the decline of information, we propose an improved DeepLabv3+ network segmentation model. Firstly, replace the backbone network with the MobileNetV2 network. Secondly, the source image is decomposed by constructing a four-channel non-separable wavelet low-pass filter, and the high-frequency subimage of the source image is extracted. Thirdly, the common convolution is replaced by deep separable convolution and the adaptive refinement feature of convolutional attention module (CBAM) is introduced to improve the segmentation effect of the network model. The experimental results show that on the VOC data set, the mean intersection over union (MIoU) of the improved model is 0.94% higher than that of the original DeepLabv3+ model, the mean pixel accuracy (MPA) is 1.34% higher than the original DeepLabv3+ model, and the accuracy is 0.19% higher than the original DeepLabv3+ model. On the BDD100K data set, mean intersection over union is 0.53% higher than the original DeepLabv3+ model. The DeepLabv3+ mean pixel accuracy is 0.15% higher than the original DeepLabv3+ model, and the accuracy is 0.13% higher than the original DeepLabv3+ model. Both subjective and objective results show that our model is better than the original model.