Fully Dilated Convolutional Networks For Semantic Segmentation

In recent years,deep learning has made greater progress in computer vision and speech recognition,especially in the direction of computer vision,from previous object recognition to object area recognition to image semantic segmentation.The image semantic segmentation problem is an image processing problem.The problem itself has a very high practical significance.It can essentially cover object recognition problems and regional detection problems.At this stage,there is a great deal of recognition in the vehicle unmanned road conditions.The help function is of farreaching significance for the future monitoring content identification and computer image understanding.The problem of image semantic segmentation is essentially the category of the object to which the pixels in the image are labeled.Most of the traditional algorithms for image segmentation are based on the degree of similarity of the pixels,or they can be divided according to the texture or the color difference between the regions,and only objects with a large difference from the background can be identified.The appearance of image semantic segmentation based on the convolutional neural network makes the image semantic segmentation problem obtain better results,but there are still problems that need to be solved.Among them,the pooling operation in the full convolutional neural network will reduce the image,and then expand the image upsampling,which will lead to the loss of image details,leading to the problem of the lower accuracy of the final result.Hole convolutional neural network is to choose to abandon the pooling layer,so as to ensure that the feature map size is consistent with the original image,and several convolutional layers are concatenated to obtain a larger convolution receptive field.However,the convolutional neural network is different from many neural network structures in image classification in the current stage.It cannot be used as a full convolutional neural network and can be directly converted from the existing network structure and trained parameters.The problem of image accuracy caused by upsampling of full convolutional neural networks,and the problem of the convolutional neural network structure caused by the change of convolutional neural network can not re-use two existing image recognition networks with better results,this article will Two key issues to solve.This paper proposes and designs an image semantic segmentation network structure called full-hole convolutional neural network.This network structure not only retains the complete structure of the convolutional neural network(reserves the pooling operation),but also ensures that it does not depend on the upsampling.Under the circumstances,by changing the original convolution layer and pooling layer to the hollow convolution layer and the hollow pooling layer,the feature map size is consistent with the original image in the calculation process,and the original convolutional neural network can be maintained.Based on the computational structure,the semantic segmentation results consistent with the size of the original image are directly obtained,thereby realizing an end-to-end image semantic segmentation algorithm in the true sense.The network structure can directly use the existing convolutional neural network for image classification as image semantic segmentation.Finally,the paper chooses to perform training validation on image semantic segmentation dataset in the MIT scene analysis challenge competition,and compares it with various methods in the dataset under the same conditions,and analyzes and verifies the feasibility and correctness of the proposed network structure.The results show that the all-hole convolutional neural network can obtain very good results and obtain more accurate results,which will play an important role in supplementing and promoting the future semantic segmentation of images.