Portrait Segmentation Based On Deep Feature Learning

Semantic segmentation is a very important research direction in the field of computer vision and pattern recognition,is widely used in medical,unmanned vehicle,and other fields.Like the traditional segmentation method based on artificial features design,single artificial extraction method,the characteristics of adaptability,the lack of robustness,and the traditional way of feature extraction is slow.With the development of deep learning,deep learning has become the most conmnon method for semantic segmentation field,having the characteristics of automatic representation and end-to-end leaming.Deep learning greatly facilitate the practical application of interoperability by the support of parallel computing platform.However,there are still many problems in the portrait segmentation task based on deep learning.In this paper,aiming at solve the problem of boundary ambiguity and other problems existing in the field of portrait segmentation,two deep learning-based semantic segmentation methods are designed to realize the efficient segmentation for portrait natural data sets.The main research contents and results of this paper are as follows.A portrait segmentation method based on deep feature integration and fusion was designed.This method aimed at the single problem of deep learning convolution feature perceptive field and the multi-scale problem of portrait segmentation,and proposed the integration learning method of multi-scale features.First by using the method of depth migration study,based on cityscapes training data set weights initialization PspNet semantic segmentation model,in the back of the features of integration,different scales have different characteristics of receptive field,namely,different models have diffeerent attention,all the characteristics of integration,further enhance the robustness of the model characteristics,the entire model of end-to-end learning and optimization.The test results show that the method is more accurate than other methods.On the basis of feature integration and fusion model in the previous chapter,a fine portrait segmentation method based on depth feature enhancement mechanism was designed to solve the boundary fuzzy problem of natural image segmentation.The method based on integration model of input output characteristics and the image fusion directly,the characteristics of the integration model of output because of a border fuzzy problem will appear similar to segmentation boundary is not fine,by the characteristics of the original image to complement and enhance the output of the integration model in the previous chapter,both designed to enhance the characteristics of a module to fusion,strengthen the characteristics of the power of expression and boundary awareness,eventually achieve the enhanced characteristics were predicted and divided.In this method,the designed enhanced feature mechanism model is tested together with the supervise data set and compared with the existing depth segmentation algorithm and the integrated depth feature fusion model proposed in the previous chapter.The experimental results show that this method is more effective.