Research On Semantic Segmentation Technology Based On Scene Analysis

As one of the very important research directions in the field of computer vision,semantic segmentation can identify the image at the pixel level,analyze the semantic category of each pixel in the image,and mark the category it belongs to,so as to make a fine scene understanding of the image.Based on the above characteristics of semantic segmentation,this technology is widely used in visual scenarios such as environmental perception,auxiliary medical imaging diagnosis,and video surveillance.However,it is difficult to realize the intensive prediction task of semantic segmentation.The main challenges currently faced can be summarized as follows:(1)Due to the limitation of training speed and memory,it is often necessary to reduce the resolution,and lowresolution images lose a lot of detailed information.(2)The upsampling method loses more detailed information,and the segmentation performance of object-like edge details and some small target categories is poor(3)The conventional convolution method is difficult to capture long-distance inter-pixel dependencies,and the global information is lost more(4)The target sizes of the same semantic category are quite different,which increases the difficulty of category determination.This paper proposes an improved and innovative method for the problem of loss of details in the process of semantic segmentation upsampling and the difficulty of multi-scale feature extraction.The work carried out in this paper can be briefly summarized as follows:(1)A novel global attention method across feature layers is proposed.In the semantic segmentation upsampling process,the skip connection method is adopted,combined with the position attention mechanism and the channel attention mechanism,a cross-feature layer global attention module is designed,and the segmentation details of the semantic segmentation network are optimized.(2)Combined with the strip pooling method,an irregular multi-scale spatial pooling module is constructed,which provides the network with an irregular perception field of view.In the multi-scale feature extraction stage,more long-distance pixel dependencies are captured,and the strip feature is provided.Greater perception.(3)Combining the global attention method across feature layers and the irregular space pooling method,a new semantic segmentation network model is designed.Under the same experimental environment,compared with Deep Labv3+,the MIo U indicator has increased by 0.77% on the Cityscapes dataset.