Research On Semantic Segmentation Algorithm Of Life Scene Images Based On Deep Learning

Semantic segmentation of life scene images is the research focus in the field of computer vision.It is a pixel-level computer vision task,and is widely used in unmanned driving,medical images,video surveillance and industrial production.Most of the current semantic segmentation models use encoder-decoder structure.Excellent encoder-decoder methods are crucial to the accuracy of image segmentation and the accuracy of segmentation edges.An excellent backbone encoding network is an important basis for semantic segmentation models,and many excellent semantic segmentation models have emerged on this basis,among which the DeepLab series network and SegFormer are representative models.However,the above models still have problems such as insufficient utilization of high and low resolution,inaccurate prediction of image segmentation edges and noise interference.To solve the above problems,this paper proposes the following two improvement strategies:1)In order to solve the problem that DeepLabV3+ does not make full use of the low-level features of the backbone in semantic segmentation and the lack of effective features caused by large factor upsampling,the cumulative distributed channel attention DeepLabV3+(CDCA-DLV3+)model is proposed.Firstly,the cumulative distributed channel attention was proposed based on the cumulative distribution function and channel attention.Then,the effective low-level features of the backbone were obtained by using the attention.Finally,the idea of feature pyramid was used for feature fusion and stepwise upsampling to output the predicted image.2)In order to solve the problem that the SegFormer model does not fully fuse high-resolution and low-resolution features and lacks multi-receptive field feature extraction for low-resolution features in the decoding stage,a conditional channel weighted spatial pyramid pooling method with less calculation is proposed.Firstly,the improved conditional channel weighting module was used to fuse the high and low resolution features more effectively.Then,the spatial pyramid pooling idea was combined with the conditional channel weighting module to obtain the features of different feelings,and then the fusion was performed.Finally,the upsampling was performed to output the model to predict the image.The experimental results show that the mIoU value of the model is higher after using the above improved method,the predicted image segmentation edge is more accurate,and the error segmentation part is significantly reduced,which proves the effectiveness of the above method.