RGB Image Semantic Segmentation Net Based on Proportional Pooling
Aiming at the problems of weak effective information of feature map and noise superposition effect of feature map noise when the traditional pyramid feature fusion segmentation algorithm performs semantic segmentation,a hybrid attention mechanisms based on proportional pooling is proposed.Firstly,the proportional pooling attention module is introduced at the feature output of the backbone network to extract the semantic information and denoise the input feature map to different degrees.and the proportion of effective feature information of the feature map is highlighted,and then the pooling results of different kernels are used as the input features of the cascade pyramid structure,and the multi-scale features after noise reduction are fused to realize the secondary feature reduction and the semantic information enhancement of small target objects.The effectiveness of the proposed method in the segmentation domain is verified on the Pascal VOC 2012 dataset,and the average pixel accuracy(mPA)and average intersection union ratio(mIoU)are used as the performance evaluation indicators of the model.Experimental results show that the pyramid network based on proportional pooling reaches 90.19％and 79.92％on mPA and mIoU,which is better than that of the comparative semantic segmentation methods.
万方数据
维普
