The Research On Matting Algorithm Based On Deep Features

Matting refers to the process of accurately extracting foreground objects in a still image or video picture sequence.Nowadays,matting technology is one of the key technologies in the field of computer vision,and it is also the basis of image synthesis technology.It has been widely used in many fields such as virtual studios,film and television productions,and image editing.However,in practical applications,it is very difficult to implement matting due to the complex and changeable images.In order to reduce the complexity of matting,the user is usually required to provide a priori information when dealing with the matting problem.The prior information is generally Trimap.Although the emergence of a trimap greatly reduces the difficulty of processing the matting problem,the accuracy of the trimap largely determines the performance of the matting algorithm.Accurate trimap often require a lot of human resources,which is obviously not desirable in real life.In addition,traditional matting algorithms rely heavily on low-level information such as the color and texture of the image,causing these algorithms to often make large errors when processing details such as hair,holes,and foreground edges in the image.Therefore,after in-depth analysis of these problems in image matting,this article starts from optimizing the prior information provided by users,and focuses on the natural image matting algorithm and its problems based on image depth feature information.The main innovations of this article include:(1)Propose a trimap optimization method based on encoder-decoder network.Because the unknown area of a rough trimap contains a large number of pixels in the front background,it affects the quality of the transparency mask image finally obtained by the matting algorithm.In order to solve this problem,this article regards it as a three-classification of pixels Using the encoder-decoder network’s ability to extract image depth feature information,and under the guidance and constraints of the cross-entropy loss function,the accurate classification of each pixel in the unknown area is finally realized,thereby solving the roughness of the three-part image.Problem,to ensure the accuracy of the three-part graph provided by the user.(2)Propose a DenseNet matting algorithm based on embedded improved SK-Net.Excessive dependence on low-level information such as color and texture in the image will inevitably affect the performance of the matting algorithm.This paper uses the DenseNet network model to fully extract the depth features,which effectively improves the shortcomings of traditional matting algorithms that rely excessively on the low-level information of the image.And the improved Sk-Net network is embedded in DenseNet,which greatly enhances the spread of useful deep feature information in the network model,and suppresses the spread of useless feature information.By using the matting results of the CF algorithm and the KNN algorithm as part of the input of the network model,the local and non-local information in the image is effectively used in collaboration.In the end,the expected effect is achieved through the training of the network model.The experimental results show that the method can effectively predict the transparency mask of the image.(3)Propose a matting algorithm based on GAN network.Deep learning-based matting algorithms improve the shortcomings of traditional matting algorithms that rely too much on low-level information of the image,but most of the matting algorithms based on deep learning ignore the multi-level connection and multi-scale fusion of deep feature information.Large errors will still occur when processing some complex images.Therefore,this paper uses the powerful generation ability of GAN network and uses nested codec network as the generator in GAN network to improve the problem.By using self-calibration convolution instead of ordinary standard convolution,the network model’s ability to extract image depth feature information is improved without significantly increasing the amount of calculation.In the end,the expected effect is achieved through the training of the network model.The final experimental results show that the method can also effectively predict the transparency mask of the image.Experiments show that the method proposed in this paper can effectively improve the defects of the existing matting algorithm,improve the quality of the transparency mask map,promote the development of matting technology,and make it more widely used in image synthesis and film and television production.