Research On Image Depth Feature Representation And Its Application

Deep learning is used to extract deep features of images from a large number of complex data.The expressive ability of deep features is better than that of shallow layer.Convolution neural network is one of the most successful models of deep learning.And it is widely used in image recognition,target detection,image restoration and so on.In this paper,we mainly studies two kinds of feature expressive networks for license plate detection and raindrop remove.It is very useful to theoretical and practical significance to promot the transformation of scientific research results into practical applications.The main contents are listed as follows:1.This paper focuses on principal component analysis network and very deep network.Principal component analysis network is used for image classification and recognition.Very deep network based on residual learning is very suitable.2.The license plate detection algorithm based on shallow features is not high in complex scenes.In this paper,a novel license plate detection method based on principal component analysis network is proposed.Firstly,the license plate candidate area is marked with Sobel operator based edge detection and edges symmetry analysis.Secondly,the deep feature extraction is performed for candidate area by principal component analysis network.And the network parameters are set by the grid search strategy;Thirdly,the support vector machine is used to confirm the license plate.Finally,we use efficient non maximum suppression to label the best license plate detection area.Experimental results show that the proposed algorithm has good robustness and high detection rate.3.Raindrops seriously affect the visual effect of the image and the application of subsequent image processing.Unlike the video rain removal,the researches of the single image rain removal have aroused the concern of domestic and foreign scholars for lacking of prior information of times and places.In this paper,a new single image rain removal algorithm based on very deep network is proposed.It can effectively use the convolution layer of the deep cascade to learn the residual characteristics between rain and rain free image brightness information.After that,we reconstruct residual features and the original raindrop images to obtain the rain free images.The method is superior to the latest raindrop image removal algorithm.