Research On Application Of Deep Learning Technology In Image Semantic Segmentation

The semantic segmentation task requires the computer to automatically recognize each category of object in the given image and assign the corresponding pixel of each category of object to the same label.Traditional image semantic segmentation algorithms use manually designed features and then construct a classifier for classification,which is difficult to achieve high performance for an advanced task.In recent years,with the rapid development of GPU technology and the emergence of large-scale data sets,deep neural network has got attention,and deep learning technology has made great progress in the field of image semantic segmentation.The deep neural network automatically learns a set of parameters to segment by minimizing the loss between image and the target.However,this method also has many problems such as how to design the decoder and how to apply it to specific problems.This paper sorts out the research work in recent years,conducts in-depth research on the problems such as feature fusion method in decoders,remote sensing small scale target segmentation and noise data,and proposes the algorithms with better performance and stronger adaptability:First,this paper investigates the feature fusion methods proposed in recent years.It designs decoders for high-level features and low-level features,introduces an attention mechanism at the back end of the decoder and proposes an enhancing feature fusion decoder.Under the two basic networks of resnet-101 and mobile net,this paper compares the decoders with different feature fusion methods.The enhanced feature fusion decoder has the best performance on the segmentation accuracy,and it is almost the same as other methods in parameter quantity and speed.Second,in this paper,the NWPU-Seg data set is produced by combining manual annotation and automatic annotation for remote sensing image segmentation.For the auto-labeling part,this paper uses the information of the target category and the bounding rectangle to segment the target using the GrabCut algorithm to obtain noise data to expand the data set.Third,for the problem of small target scale and over-fitting in remote sensing image segmentation tasks,this paper proposes a high-resolution segmentation network,which has a lower downsampling multiple and a smaller receptive field,and has better adaptability to the task.And then,this paper improves the weakly supervised image classification network and proposes a dual-decoder image segmentation network based on a high-resolution network.The two-branch structure of the two decoders of the network is trained using manually fine-labeled data and automatically generated data with a large number of incorrect labels,so that the network learns more robust parameters and improves segmentation accuracy.