Research On Image Semantic Segmentation Algorithm Based On PSPNe

Image semantic segmentation is the process of assigning a semantic label to each pixel in an image,enabling accurate recognition of different objects and backgrounds.Traditional methods typically use manually designed feature extractors and classifiers,which are difficult to provide high-precision results when dealing with complex scenes and diverse datasets,and cannot adaptively process different image scenes and datasets.In recent years,with the development of deep learning technology,image semantic segmentation methods based on convolutional neural networks have made great progress.This paper focuses on the study of image semantic segmentation based on PSPNet,and proposes two new image semantic segmentation algorithms from different perspectives to improve PSPNet for poor segmentation results caused by factors such as object shape,category distribution,and object occlusion.Firstly,to solve the problem of reduced segmentation performance due to irregular object shapes and uneven category distributions,a context-aware SA-PSPNet image semantic segmentation algorithm is proposed.By using a long strip-shaped window in the bar-pooling module for pooling operations,the network can perceive a larger range of image areas and better identify and segment targets.At the same time,the barpooling module can consider features in different directions to better perceive information about the edges of the target.SA-PSPNet uses the Lovasz-Softmax loss function to consider the contribution of each pixel’s classification error to the loss,rather than simply calculating the classification error rate for each category.This allows it to better handle situations where the number of pixels is uneven,thereby improving segmentation accuracy for less frequent pixel categories.Secondly,to solve the problems of unclear object edges,small object shapes,and errors caused by object occlusion,an attention-based AM-PSPNet image semantic segmentation algorithm is proposed.In the encoding phase,an efficient channel attention module is used to extract basic residual network features,learning the importance weights of each channel and allowing the network to pay more attention to channels with category information,while ignoring channels that are irrelevant to the segmentation task,thus improving the network’s accuracy and generalization ability.AM-PSPNet uses a deep guidance fusion module in the decoding phase to fuse deep features with effective shallow features,retaining more critical detail information and high-frequency features,and improving the network’s ability to process detail information.Finally,this study evaluates the performance of SA-PSPNet and AM-PSPNet on the Cityscapes dataset,PASCAL VOC 2012 dataset,and Vaihingen dataset.Compared to the original PSPNet’s m Io U values,SA-PSPNet improves by 1.5%,1.2%,and 1.6%for the three datasets,while AM-PSPNet improves by 1.7%,1.5%,and 1.8% for the three datasets.The experimental results demonstrate that the proposed networks,SAPSPNet and AM-PSPNet,outperform existing semantic segmentation networks such as FCN-8s,U-Net,Deep Lab V3,and PSPNet on different datasets,effectively enhancing the segmentation performance of the networks.