Research And Application Of Image Features Based On Neural Networks

With the development of computer vision technology in recent years,the demand for image feature extraction,analysis and processing has also increased.Image feature extraction is the basis of advanced computer vision tasks such as image classification and recognition.Appropriate image feature processing methods can effectively improve the expression ability of image features,which means that the obtained image feature information is more accurate.Therefore,research on image feature processing methods is necessary.Scale Invariant Feature Transform(hereinafter SIFT)is a general-purpose image feature extracted by traditional hand extraction algorithm.SIFT can remain invariant in operations such as rotation,scaling and brightness changes,thus has a wide range of applications in image matching and recognition.Neural network is a current typical algorithm for pattern recognition,and it performs well in multi-classification problems.The main research of this thesis is to process the SIFT features extracted from the image combined by the neural network,and then apply this processing method to the recognition task of the parts on poles and towers.The main contributions of this thesis is as follows:1.There are differences in the number of SIFT features that can be extracted by different images,which brings many challenges to subsequent processing.Aiming at this problem,this thesis summarizes the common coding methods for solving this problem,and proposes a new coding method based on local aggregation vector(hereinafter VLAD),which solves the local feature information loss problem in the original method.Because the SIFT features lost the description of color,the BP neural network is built to fusing the color features and SIFT features.And our way has better performance in the recognition experiment of the parts on poles and towers than the other.algorithms.2.In the related research and application of SIFT features,only the feature vectors of SIFT features are usually used,and the correlation of SIFT feature points is not well analyzed and used.Based on this problem,this thesis proposes an algorithm for constructing a graph structure to describe the local feature set of the original image by combining the correlation information of the SIFT feature points and the SIFT feature descriptor information.After that,this thesis designs a convolutional neural network for processing constructed graph based on the theory of graph vertex domain.In the recognition experiment of the parts on pole and tower,the processing method which adds the correlation information of SIFT feature points has better recognition effect than the previous method.3.Based on the proposed algorithm,this thesis designs and implements an image SIFT feature analysis system based on neural network,completes the related coding and system testing work,and shows the running results of the system in detail.