Content Based Image Feature Extraction And Application

With the rapid development of mobile Internet technology and digital storage and communication devices,hundreds of millions of images are transmitted by the social networking applications and instant messaging applications.Especially in the face of such huge amounts of image data,how to retrieve the required data quickly and accurately,and how to extract structured,meaningful logic,entities,and even relational networks from these unstructured data,so as to predict the social groups'trends and individual behaviors is a new challenge and new opportunity in the era of Internet big data.By extracting features that are strongly correlated with images,not only the data storage,transmission,and learning costs can be reduced,but also images can be portrayed more deeply to achieve efficient image matching and retrieval.Based on the theories of computer vision,information coding,complex network and optimization,the research takes the feature extraction algorithm and it's application as the research object,and we focus on the extraction of edge orientation features,composite features and interest points features.We focus on how to construct a good image representation,how to describe the image content more completely and how to perform effective real-time assessment of image retrieval results.The main innovations in this article are summarized as follows:(1)For the problem of "edge degradation",we firstly proposed an image compression coding and feature extraction algorithm based on classified Vector Quantization(VQ)and the improved edge orientation classification template.An edge direction detector based on Edge Orientation Classification Patterns(EOPs)is proposed,and we apply the edge orientation detector to classify the image blocks into smooth block and direction blocks,and design the sub-codebook for different blocks by using the classified vector quantizer.Then the input vectors are replaced by best matching indices of the codebook,and the indices are transmitted and stored to gaining high coding efficiency with high average PSNR.In addition,we proposed a new fused feature based on the EOPs and VQ,combining the EOP-histogram(EOPH)-based features with the traditional VQIHs.In the experiment,the experimental results show that our scheme effectively improves the image retrieval performance.(2)Aiming at the difference of local feature in different regions of the image,the composite features schemes are proposed to overcome the problem of local information loss.As Block Truncation Coding(BTC)is known for its high quality of reconstructed images,much lower computational complexity and low compression ratios,we propose a hybrid feature which merges BTC and Tree-Structured Vector Quantization(TSVQ)with the classification of edge orientation bit-planes.The bit-plane is gradient-classified into active and inactive blocks according and the active blocks containing edges are encoded with BTC and EOBP templates,and the inactive blocks having low intensity variations between its pixels are encoded with the VQ.Finally,we obtained the Edge Orientation of Bit-plane Feature(EOBPF)and the Contrast and Colour Histogram Based Feature(CCHF).Simulation results show that our proposed feature not only has a high compression bit rate,but also effectively overcomes the problem of partial information loss and greatly improves the efficiency of image retrieval.(3)As most of the images and videos are stored and transmitted in compressed form,we propose a feature extraction method based on discrete cosine transform in the compressed domain.An image retrieval scheme,called Classified DCT based Vector Quantisation Index Histogram(CDCT-VQIH)is proposed to extract features from the DCT frequency domain.Our algorithm takes advantage of both the energy-compaction property of DCT and the high compression ratio of classified vector quantization(CVQ).Our CDCT-VQIH based feature combine the class index,the DC index and the VQ index together,which represent the edge classification information,energy information and texture information of an image respectively.The retrieval simulation results show that,the algorithm implements the composite feature extraction of compressed domain images and removes irrelevant information,with performs much better in terms of recall and precision and has practical application value.(4)To avoid "visual degradation" issues,we studied the complex network-based image modeling algorithm and complex network-based image point feature extraction method,and proved the invariability of the feature.Firstly,the image is mapped into a full-connected graph through the pixel points.Then the full-connected graph is dynamically evolved to generate a complex network with obvious topology.Then,we use a multi-neighbourhood gradient map of the target image to build a network.In this step,we use a set of thresholds and a Gaussian weighted function to reduce the complexity of the generated network as well as keep the local properties of the original image.According to the structural characteristics and topological characteristics of complex networks,we find the critical nodes and links based on the entropy method and information entropy and cross-information entropy in complex network.Then the critical nodes are mapped into the image,and the feature of interest points capable of describing the shape contour of the image is obtained.After finding the feature points,we take a "range" around the feature points and calculate the statistical characteristics of this small sub-network as a local feature descriptor of the image.Simulation experiments show that the image feature points we extractd not only describe the local features of the image well,but also have scale and zoom invariance,which is in favour of image analysis and understanding.