Research On Image Segmentation And Matting Based On Deep Learning

Image segmentation is a research hotspot in computer vision,and its application areas include medical image analysis,scene understanding,robot vision,virtual reality,film and television special effects,etc.With the improvement of computer hardware,image segmentation based on deep learning will gradually emerge in these fields.In response to the above image segmentation problems,this paper proposes image segmentation and matting operations based on deep learning.This paper first improves a spectral clustering algorithm based on convolutional neural network,which combines deep learning and traditional image segmentation algorithms,and uses pre-trained convolutional neural networks to perform feature extraction and feature fusion on images.The feature space of the original image is effectively constructed,thereby reducing the dependence on the similarity matrix.In the spectral decomposition process of the similarity matrix,the Nystrom approximation method is used to approximate the feature space of the similarity matrix,thereby accelerating the speed of image segmentation.Finally,the Berkeley image data set proves that the algorithm can effectively reduce the memory and time consumption of spectral clustering,and improve the accuracy of image segmentation.In order to obtain higher-quality image segmentation,this article has improved a fully automatic matting algorithm based on deep learning.This algorithm mainly addresses the problems of low character matting in matting tasks,insufficient edge refinement,and cumbersome matting tasks.The algorithm uses a three-branch network for learning,the semantic information of the SSB branch learning ?,the detailed information of the DB branch learning ?,and the COM branch summarizes the learning results of the two branches.First,the algorithm's coding network uses a lightweight convolutional neural network MobileNetV2 to speed up the algorithm's feature extraction process;secondly,an attention mechanism is added to the SSB branch to weight the importance of image feature channels,and a hollow space pyramid is added to the DB branch The pooling module performs multi-scale fusion of the features extracted from different receptive fields of the image;then the two branches of the decoding network merge the features extracted by the different stages of the encoding network through the jump connection to decode,and finally merge the features learned by the two branches get the ? of the image together.Experimental results show that this algorithm is better than the traditional matting algorithm and the matting algorithm based on deep learning on the public data set,and the effect of matting in real-time streaming video is better than Modnet.