Research On Image Semantic Segmentation Based On Deep Learning

Image semantic segmentation is the classification of all pixels in an image.It belongs to the pixel level classification.With the development of deep learning,more and more studies use convolutional neural networks to identify and segment complex images,and they are widely used in areas such as autonomous driving,medical imaging,geographic remote sensing,etc.and show good performance.However,due to the existence of the downsampling operation in the convolutional neural network,the objects in the segmentation result will lose more detailed information and the edge of the segmentation category will be blurred.In addition,the current semantic segmentation model in order to improve the accuracy of the segmentation results in a high model calculation and long training time,which makes it difficult to adapt to a variety of environments.Therefore,this paper studies the existing image semantic segmentation algorithm,from the point of view of detail information processing,segmentation edge,multi-scale object recognition and calculation in the process of segmentation,designs an improved algorithm with accurate segmentation and high efficiency,and verify the effectiveness of the algorithm through experiments.The main research contents are as follows:First,to solve the problem of rough segmentation of category edge in current semantic segmentation,this paper combines the residual module and attention mechanism,designs the residual attention module based on the structure of encoding and decoding,and we propose the encoding and decoding network based on the residual attention mechanism.we focus on segmenting the edge of categories by attention mechanism,then strengthen the connection between the channels of the feature map.Finally,it is combined with the residual module to extract the deep semantic features and pay attention to the image details,then the accuracy of semantic segmentation has been improved.Secondly,in order to reduce the computational complexity and amount of computation caused by ordinary convolution,we introduce the deep separable convolution and replace the ordinary convolution of the coding part with the deep separable convolution,so that the segmentation performance of the model can be guaranteed while reducing the amount of calculation.In order to solve the problem that multi-scale objects cannot be segmented effectively in semantic segmentation,this paper improves the decoder part and designs a joint pyramid sampling module.The module extracts semantic information of different scales by using different scale feature maps generated in the coding process,so that we can improve the recognition ability of the model to multi-scale objects,we also integrate the feature pyramid to enhance the understanding ability of the model to the scene.Finally,in order to verify the effectiveness of the proposed network structure,we perform experimental verification on Pascal VOC 2012 and Cityscape datasets.We compare the method proposed in this paper with the more advanced method.Experimental results show that this method has a great improvement in segmentation accuracy,and the segmentation effect is also better.