Research Of Image Semantic Segmentation Based On Deep Convolutional Neural Network And Boundary Optimization

Image semantic segmentation is a hot research field of computer vision and has been widely used for video detection,automatic driving,human-computer interaction and many other applications.With the application of deep learning in various tasks of computer vision,researchers have found that deep convolutional network can achieve excellent results in image semantic segmentation,which is attributed to the rich features extracted from various convolution kernels in the convolutional network.The common deconvolution operation in deep convolutional network makes the prediction accuracy of object boundary in the final segmentation result lower.To improve the accuracy of semantic segmentation,we propose deep convolutional network based on clustering algorithm and deep convolutional network based on boundary mapping for the disadvantage of low boundary segmentation accuracy of deep convolutional networks.We verify the effectiveness of the proposed method on the semantic segmentation database.The main work and contribution of this work are: 1)We propose a deep convolutional network based on clustering algorithm.The architecture of this network is parallel.Deep convolutional networks have powerful ability to extract object features,while clustering algorithms can obtain clear object boundaries.In order to combine the advantages of the two methods,we define a new score function which can use the prediction of clustering algorithm to optimize the boundary segmented by deep convolutional network.The accuracy of semantic segmentation is improved in this way.Parallel structure makes the algorithm have strong generalization ability.Many deep convolutional networks and algorithms can be applied in this structure.We select three different deep convolutional networks and conduct experiments on two standard datasets: PASCAL VOC 2012 and Cityscapes.The experiments demonstrate the effectiveness of this approach.The noise experiments verify the robustness of this method in the case of noise.2)We propose a deep convolutional network based on boundary mapping.We design a boundary refinement block and add it to the deep convolutional network.The block extracts boundary feature maps from the feature maps extracted by the deep convolution network and outputs the segmentation result of the object boundary.The deep convolutional network obtains the preliminary segmentation result,and the boundary refinement module obtains the segmentation result of the object boundary.The final result is obtained by adding the two partial results and then normalizing it.The segmentation result of the object boundary is optimized in this way,thereby improving the accuracy of semantic segmentation.We rebuild the ground truth which is used for the training of the block.We define a new loss function and introduce the training details.The experiments on PASCAL VOC 2012 and Cityscapes achieve good results.This method not only improves the segmentation accuracy of object boundary,but also makes the prediction time less than the deep convolutional network based on clustering algorithm.