Research On Image Semantic Segmentation Based On Feature Enhancement Neural Network

With the rapid development of automation and intelligence technologies,Image Semantic Segmentation as an important computer vision technology,has been applied to medical image processing,autonomous driving,aerospace and many other fields.Traditional image segmentation methods can only perform the extraction of simple image semantic,such as distinguishing targets and backgrounds,cutting target boundaries,while complex application scenarios require the information about target positioning and classification.With the rise of deep learning methods,a large number of method using convolutional neural network take over the semantic segmentation completed by traditional methods,classifying and labeling the pixels in the different semantic regions of image.In complex scenes,there are many types of targets with large changes of scale.Nonsignificant targets with small scales might be incompletely segmented or missed,which is a problem for the refinement of semantic segmentation.The semantic segmentation using convolutional neural network is facing the process of image resolution reconstruction,where simple fusion of feature maps leads to confusion of pixel segmentation between targets on a larger scale.In view of above problems,this thesis proposes a feature enhanced U-shaped convolutional neural network(FEUNet),which is based on a U-shaped network as known as encoder-decoder framework for image feature extraction and resolution reconstruction,the main contributions are:(1)Designed a Local Feature Enhancing Module(LFE)at the encoding stage to highlight the local perceptive features used to segment smaller-scale targets through dilated convolution,and fine the coarse or neglected segmentation of non-significant targets at small scales.(2)Designed a Global Feature Enhancing Module(GFE)during the decoding phase,which uses global pooling operations to extract the global context information in each of deep layer feature channels,to enhance the context information of shallow layer features,and narrow the difference between the two parts to realize a better fusion,the GFE solves the problem of insufficient discriminative features among targets leading to a confused segmentation.The experiment results on public datasets prove the effectiveness of the proposed module,and the superiority of FEUNet compared with other existing segmentation methods.In order to improve the segmentation performance of the module,this thesis introduces a Weakly Supervised Learning Branch based on the proposed method FEUNet,supervising the basic network to extract all target features.The class label of the targets in the image is used to implement the image multi-label learning task,assisting in the final semantic segmentation task.The weak or semi-supervised learning of the model is achieved by the branch in case of insufficient segmentation labels.In addition,before the final output of the decoder,the Densely Adjacent Prediction is added to rich spatial information of output by expanding the feature map channels,resulting in a better resolution reconstruction with a small number of parameters increase.The experiment results on public dataset show that the performance of the proposed method has been further improved.