Research And Application Of Deep Learning For Portrait Segmentation

Computer vision has always been a hot topic of research,and its research results are widely used in all aspects of social life.With the explosive development of the deep leaning in recent years,the research and application of deep learning in computer vision has made significant progress in various tasks in computer vision.From the traditional method based on simple statistic to the deep learning method based on pattern recognition,computer vision has achieved a staged breakthrough.Image segmentation is one of the most important tasks in computer vision.Its research progress also represents the latest achievement of computer vision.Based on the research status of image segmentation in deep learning,this paper analyzes the effectiveness,the applicable problems of existing methods.And the problems are explained of existing methods and the corresponding improvement methods are proposed in this paper.The main research contents of this paper are as follows:1.The fully convolutional networks(FCN)for the image segmentation have the disadvantages of the noise,boundary roughness and no prior shape.A level set segmentation method based on shape prior and deep learning for image segmentation is proposed in this paper.The output of the FCN is treated as a probability map in the proposed method,and the global affine transformation method is used to transform the inherent object shape to a corrected shape corresponding to a specific image.Finally,the improved level set method is used to integrate the information of the original image,the probability map and the corrected prior shape to achieve the image segmentation.2.The image can be regraded as a two-dimensional signal.Based on the wide application of frequency domain-based methods in signal analysis,a multi-frequency decomposition-based neural network method is proposed in this paper.Based on the fast Fourier transform,the multi-frequency decomposition is made into a layer of neural network,which can be integrated into any network structure.Depending on the ability of feature extraction of deep neural networks,multi-frequency decomposition layer can extract effective information of different frequencies in images.3.The result of existing neural network learning is not the logic of human recognition,but tend to remember some easily distinguishable features in images.A neural network structure based on attention learning is proposed in this paper.The proposed method attaches additional attention constraints to the neural network,making its learning results more inclined to human recognition logic.4.A portrait segmentation system based on web microservices is designed and implemented in this paper.