Retrieval Research On Image Semantic Segmentation Algorithm Based On Deep Learning

With the development of deep learning technology,more and more problems in the field of computer vision need to be solved by deep learning networks,such as image semantic segmentation.The application of deep convolutional neural networks to the field of semantic segmentation requires three challenges: First,the convolution operation with a step size greater than one and the existence of the pooling layer can greatly help feature extraction,but also bring images.The problem of reduced resolution and loss of spatial information;second,in the upsampling operation of restoring the original image size,the general network only utilizes the characteristics of the deep network and a small number of shallow networks,which may result in the restoration of the image due to the lack of low-level information.The positioning is not accurate or the edges are blurred.Thirdly,most networks use deep neural networks with large parameters due to the precision of image semantic segmentation,which makes the network runtime very slow.How to ensure its real-time performance is also one of the challenges.This article addresses these issues and makes some improvements based on existing work:(1)In order to make full use of all the useful information contained in low-level and high-level features,this paper designs and uses two parallel networks to extract the positioning information contained in the lower level and the abstract feature information in the high-level classification..A "residual block" structure similar to the ResNet deep learning network is used in each network to fully integrate the information of each layer.Finally,the two parallel networks are upsampled at the same scale and then network fused,and the fused feature map is upsampled to restore the original image size,thereby completing the image semantic segmentation task.Based on these improvements,good results have been achieved on the Street View dataset.(2)For the problem of image semantic segmentation,the classification task and the localization task are inherently contradictory.The full convolutional neural network with global convolutional layer and fully connected layer is used to perform semantic segmentation on the selection of convolution kernel.In order to achieve a better classification task,unlike other networks that use small convolution kernel superposition instead of large convolution kernel,a large convolution kernel is used to extract a larger receptive field covering almost the whole picture,even in the image.The scale of the object varies greatly,and the network can successfully classify it.Considering the large amount of large convolution kernel parameters,it is proposed to decompose a large k×k convolution kernel into two combinations of 1×k and k×1,and the relu activation function is not connected in the middle,thus ensuring a large The effect of Kernal reduces the amount of parameters.Compared with the traditional FCN network,the effect is significantly improved.(3)Combine the structure of the first two networks and the selection method of the convolution kernel,and replace the backbone network with a lightweight deep learning neural network to realize real-time semantic segmentation on the autopilot data set CamVid.