Research On Image Semantic Segmentation Algorithm Based On Deep Learning

Semantic segmentation is a kind of typical computer vision problem,its essence is the pixel level of an intensive classification task.In recent years,with the rapid development of artificial intelligence technology,semantic segmentation technology based on deep learning has been widely used in scene understanding,medical image analysis,robot perception,video surveillance,augmented reality,and image compression.However,due to the diversity and complexity of the shape and color of objects in the real world,the current semantic segmentation algorithm based on deep learning has some shortcomings in the multi-scale problem of segmentation target,target occlusion problem,and small target recognition problem.Because of the above research difficulties,this paper proposes corresponding improved methods,which can achieve accurate detection and positioning of complex targets.It can effectively improve the performance of segmentation and obtain more accurate segmentation results.The main research contents of this paper are as follows:For the single object semantic segmentation scenario,this paper proposes a feature recovery module based on asymmetric encoder-decoder architecture and a multi-scale feature extraction module based on dilated convolution.The feature recovery module based on the asymmetric encoder-decoder architecture effectively reduces the information loss in the down-sampling process by reducing the depth of the decoder.The multi-scale feature extraction module based on dilated convolution effectively makes up for the problem of error accumulation in the decode architecture by increasing the connection between the encode and the decoder and improving the segmentation effect of the segmentation algorithm in the single target scene.Finally,a single target semantic segmentation network with bidirectional channels is constructed by combining the above modules.Experimental results show that the proposed algorithm can extract features of different scales better and achieve good segmentation performance at a lower computational cost.For multi-target semantic segmentation,a feature extraction method based on a linear attention mechanism and a feature fusion method based on a cross-channel attention algorithm is proposed.The feature extraction method based on the linear attention algorithm can obtain a wider range of global information while reducing the complexity of the self-attention mechanism from O(n~2)to O(n).The feature fusion method based on the cross-channel attention algorithm uses high-level semantic features and low-level spatial features to calculate the global attention matrix and fuses the global attention features and low-level spatial features,which makes up for the lack of spatial details in the semantic segmentation network and improves the segmentation accuracy of complex targets.Finally,a real-time semantic segmentation network based on a single channel algorithm is proposed.The experimental results show that the network can optimize the segmentation accuracy and improve the reasoning speed in many complex target semantic segmentation tasks.