Research Of Visual Semantic Segmentation Technology Based On Convolutional Neural Network

Semantic segmentation is one of the key issues in the field of computer vision.Its goal is to classify data such as images and videos at the pixel level,and give each pixel a label.This is a task applied to scene understanding.Because of its broad application prospects in medical diagnosis,autonomous driving,robotics,augmented reality,etc.,visual semantic segmentation technology has become a hot research topic.With the many breakthrough achievements of deep learning in the field of computer vision,a large number of deep learning methods are used in semantic segmentation.Based on deep learning theory,this thesis studies visual semantic segmentation technology based on convolutional neural networks,including semantic segmentation of images and videos.The main work of this thesis is summarized as follows:1.Aiming at the problems of insufficient feature extraction,insufficient feature utilization and large amount of parameters in semantic segmentation,a multi-feature fusion lightweight image semantic segmentation network is proposed.The network is based on the ”encoder-decoder” structure.In the ”encoder”,the network combines deep separable convolution with dense connections,reducing the number of network parameters,making the network structure lighter,and simultaneously connecting the characteristics of different network layers to enhance the ability of the network to extract features;in the ”decoder”,the network uses an optimized feature fusion method to fuse the features extracted by the ”encoder” with the corresponding features in the ”decoder”,improving the accuracy of the network.During the training process,the network uses a weighted loss function,which can make the loss function converge faster,reduce the training time,and also help to improve the accuracy.Taking medical aided diagnosis as the application background,the multi-feature fusion lightweight image semantic segmentation network is verified on the retinal fundus image dataset DRIVE,and the best segmentation effect has been achieved.2.Aiming at the problems of complex calculation,slow running speed and difficulty in real-time video processing in semantic segmentation,a video semantic segmentation network combined with optical flow is proposed.Video is different from a single image,and there are a lot of spatiotemporal correlations between video frames.The proposed network can effectively use the correlation between video frames to achieve the purpose of reducing the amount of network operations and increasing the speed of network operation.In the process of processing video,the network performs image semantic segmentation on some sparsely distributed video frames,and the remaining frames obtain their semantic segmentation results by transmitting features through optical flow.This method greatly improves the network's processing speed.Taking automatic driving assistance as the application background,the video semantic segmentation network combined with optical flow has been verified on the urban streetscape datasets Cityscapes and CamVid,and has achieved leading performance.3.Aiming at the problem of better selecting key frames in video semantic segmentation,an adaptive video semantic segmentation network is proposed.In the process of video processing,it is necessary to perform image semantic segmentation on some sparse frames.These sparse frames are key frames,and the selection of key frames is an important issue.In order to better select key frames,the network proposes an adaptive key frame selection strategy,which can adaptively determine key frames according to scene changes.The strategy can achieve better balance between processing speed and segmentation accuracy.Taking automatic driving assistance as the application background,the adaptive video semantic segmentation network has been verified on the urban streetscape datasets Cityscapes and CamVid,and has achieved leading performance.