Optimization Of Human-Object Semantic Segmentation Algorithm Based On Feature Aggregation

Understanding the semantics of people and objects in images is a difficult point in the field of artificial intelligence.The main task is to use deep learning,reinforcement learning and other algorithms to integrate serialized image feature information to determine the areas of people and objects in the image.In recent years,the development of traditional machine learning semantic segmentation technology can basically meet the needs of locating people in images,but deep learningbased semantic segmentation still needs to be studied.The research goal of this dissertation is to identify each semantic in an image.First,a semantic feature extraction method based on attention mechanism is designed to extract the multi-dimensional fusion features of the main content of the image,and then an edge information enhancement algorithm is designed to reduce the boundary information during network training.loss,and finally the wrong pixel re-labeling training module re-labels the pixels that do not meet the requirements.The innovation of this thesis is mainly reflected in the following three aspects:(1)Use the fractional differential edge detection operator to generate an edge information map,fuse the edge information of the original image,strengthen the boundary information between the various semantics of the feature map,and reduce the feature loss of the boundary information in the process of network training.Semantic segmentation and ablation experiments are carried out on the City Scape dataset.The experimental results show that the boundary feature enhancement algorithm based on fractional differential operator designed in this thesis can reduce the distortion rate of boundary information during network training and retain the boundary semantic feature information of the original image.(2)Using the spatial attention mechanism and the channel attention mechanism,the network acts on the feature layer extracted by the backbone network,strengthens the feature representation ability,and fits the original image and the deep feature map through the attention mechanism,so that the network only focuses on specific semantics part to improve segmentation accuracy.Semantic segmentation experiments are conducted on the City Scape dataset.The experimental results show that the feature extraction algorithm based on the spatial attention mechanism designed in this thesis achieves excellent segmentation results and can better locate the target area.(3)Re-label training is carried out by using the wrong pixel re-label training method,and the predicted image is regenerated through the error network module and the detail network module,and a binary classification model of the original image misclassified label is constructed.Semantic segmentation experiments are carried out on the City Scape dataset,and the experimental results show that the error-pixel-based re-label training algorithm designed in this thesis can improve the robustness of the semantic segmentation network algorithm.