| With the rapid development of the Internet and its wide application in various aspects of real life,a large amount of image data has been generated.There are many uncertainties in the complex environment of real life,such as many kinds of targets,obscured objects,and uneven illumination,resulting in a lot of room for improvement in the accuracy of image semantic segmentation.In recent years,the image segmentation method based on deep learning has achieved great success,and the technique of image semantic segmentation can assist the computer to better understand the content expressed in the image,whose main task is to filter whether a given image includes a certain type of target object,mark the semantic category of each pixel in the image,and draw the boundary range of each object,and finally get the segmented image with the semantic identifier of the pixel.The final segmented image with pixel semantic identifiability is obtained.In the semantic segmentation task of images,due to the differences in pixel values between different objects,this paper conducts an in-depth study of image semantic segmentation methods,and the research points in this paper are as follows.(1)Aiming at the problems of invalid feature information generated in the process of image semantic segmentation and the loss of local detail information of images,the paper proposes an encoder based on DCNN,ECANet,and DSA＿ASPP.The DCNN in the model extracts feature information of the image;ECANet improves the network representation capability of the model,enhances channel features,and suppresses useless information;DSA＿ASPP performs multi-scale processing and fusion of feature maps to achieve efficient capture of local detail information.(2)Based on the above encoder,a DECANet network model is proposed,and the pixel-level information contained in the extracted feature maps is more accurate.Finally,the Mean Intersection Over Union(MIo U)reaches 81.08%,79.73%,and 76%on the PASCAL VOC2012,Cam Vid,and Cityscapes datasets,respectively,compared with the existing advanced network models,the method in this paper has a detection performance is better,which can effectively capture local detail information and reduce image semantic pixel classification errors.(3)In the scene recognition task,the main objective is to identify various object objects contained in the scene,and then segment them to capture information such as the outline of the object and the category to which it belongs,to provide technical assistance in the subsequent scene recognition tasks.In order to reflect the application value of the algorithm proposed in this paper,a encoder-decoder structure-based application tool for street scene recognition is designed in this paper,which locates the specific location and category attributes of the target objects by semantic segmentation and visualizes the recognition results. |
