Semantic Image Segmentation Method Based On Deep Learning

Image segmentation is a kind of technical process that divides the image into different characteristics and does not overlap each other.It is the most basic preprocessing means in artificial intelligence and computer vision.It is one of the most important technical branches in AI and autopilot.The quality of image segmentation will directly determine the accuracy of image fusion and image recognition.Traditional image segmentation methods for complex background image segmentation accuracy is limited,it is difficult to accurately and quickly achieve pixel level image segmentation.The continuous pooling and downsampling in the traditional full convolutional network FCN will reduce the resolution of the feature map,which makes it difficult to achieve accurate classification of image pixel semantic labels.To solve the problem of insufficient segmentation accuracy in traditional semantic segmentation,this paper combines the improved convolutional neural network deeplav3 + with super-pixel algorithm,and studies the semantic image segmentation method based on deep learning.Aiming at the problem of the boundary adhesion of weak edges and the regular superpixel shape in the traditional SLIC(Simple Linear Iterative Clustering)algorithm,an improved Canny SLIC image segmentation algorithm based on gradient direction is proposed,namely "C-SLIC Algorithm".The improved Canny algorithm is combined with SLIC by using the difference between the edge and noise in the gradient direction,and then uses complex number operations to reduce the dimension of image edge features,and finally uses hexagons to describe the superpixels generated by SLIC.Experiments show that the algorithm can The superpixel segmentation map that closely fits the object boundary and has high segmentation accuracy is generated,which verifies the effectiveness of the C-SLIC algorithm.Aiming at the problem of insufficient segmentation accuracy in the traditional Deeplabv3+ network model when the training period is short,the Deeplabv3+ network model is improved and the "M-deeplab model" is obtained.In the Deeplabv3+ network model,the lightweight attention module CBAM(Convolutional Block Attention Module)is introduced to convolve the attention map with the input feature map to achieve adaptive feature optimization,thereby increasing the cardinality of the neural network model.Improve the ability of neural network to extract features.In order to further improve the segmentation accuracy of the M-deeplab model on the target edge area,the M-deeplab model and the C-SLIC algorithm are combined.The C-SLIC algorithm is used to obtain the superpixel segmentation boundary information of the image,and based on the boundary information,new superpixel semantic annotation criteria are introduced to improve the sensitivity of the model to the target edge area,thereby optimizing the M-deeplab Based on the semantic segmentation results obtained by the model,a "CM-deeplab model" is established.The experimental results show that the overall average cross entropy Miou and pixel accuracy PA of cm deep lab model in this paper are 71.8% and 92.6% respectively on the latest Pascal VOC data set,and the segmentation accuracy is better than the other compared.The model and the semantic segmentation map are also improved visually,which verifies the effectiveness of the CM-deeplab model.