Research On Image Semantic Segmentation Method Based On Improved ASPP

The powerful feature extraction capabilities of convolutional neural networks have made an important contribution to the development of computer vision tasks.The main goal of the semantic segmentation task,one of the fundamental branches,is to interrelate each pixel point in an image with a certain class,thus dividing the image into different regions.Semantic segmentation techniques are currently widely used in areas such as remote sensing measurements,medical image segmentation and human body analysis.Existing segmentation methods focus on improving semantic segmentation accuracy,real-time and applicability,but there are still some methods that lack effective utilisation of the extracted image features.In order to improve feature utilisation and the accuracy of existing semantic segmentation methods,the following two improvement methods are proposed:(1)Building an improved ASPP and multi-level feature semantic fusion segmentation method(MFSF),which adopts an encoder-decoder structure as a whole.The encoding part uses Res Net50 as the feature extraction network,and the decoding part groups the extracted features equally by channel,and uses the Split Atrous Spatial Pyramid Pooling(SASPP)module to obtain the multi-scale contextual information of each group,and introduces the Strip Pooling(Strip)branch to improve the SASPP module’s performance on the segmentation.The SASPP module can improve the ability to capture contextual information in striped regions.The Semantic Guidance Fusion Module(SGFM)is proposed to fuse the global semantic information with the local detail information in order to solve the problem of the difference in the corresponding pixel positions when fusing features at different levels.The experimental results show that the MFSF segmentation method achieves 73.1% and 71.8% segmentation accuracy on Pascal Voc 2012 and Cityscapes datasets respectively,which proves the better semantic segmentation performance of the MFSF method.(2)The improved segmentation method Deeplabv3++ was proposed to further improve the segmentation accuracy of Deeplabv3 method.The Spatial Attention Module(SAM)is embedded in the ASPP module to improve its ability to selectively aggregate multi-scale contextual information,making the aggregated features more specific to the semantic segmentation task.The Channel Attention Module(CAM)is used to weight the aggregated features by channel to obtain interdependency information between different channels.The Feature Fusion Module(FFM)is used to integrate the shallow detail information into the global contextual semantic information to optimise the target detail segmentation results.The experimental results show that the segmentation accuracy of Deeplabv3++ is 3.1% and 1.6% higher than that of Deeplabv3 and MFSF methods respectively on Pascal Voc 2012 dataset,and 2.6% higher than that of Deeplabv3 method on Cityscapes dataset,demonstrating the effectiveness and rationality of the improved method.