Research On The Description And Selection Of Color And Texture Features For Image Retrieval

With the rapid development of Internet and multimedia technology,a large number of data in the form of audio,video and image are being applied in many fields such as medical care,satellite data,video,still image storage,digital forensics,and surveillance systems.It leads to a continuous demand for a system that can efficiently store and retrieval multimedia data.Image retrieval is defined as the retrieval of semantically related images from an image database.The automatic derivation of semantically meaningful information from the content of images is the focus of most image database research.As this point,the content based image retrieval has been proposed.In order to meet these demands,researchers have developed a huge number of multimedia information storage and retrieval systems.This paper mainly focuses on how to design an efficient content-based image retrieval system.The emphasis of the design is on the extraction and description of the underlying features,the integration and selection of multiple features,etc.CBIR(Content-Based Image Retrieval)automatically extracts low-level features such as color,texture,shape and spatial location from images to represent images in search database.Then the retrieval system is dynamically adjusted according to the user's feedback mechanism.In many methods,some of the underlying features are fused into a high dimensional feature,then the retrieval task is performed,and the satisfactory results are obtained.However,the high dimensional features of this method often have problems such as redundant information,and even noise.This not only increases the calculation time,but also may have a series of problems such as overfitting,low efficiency,poor learning performance and so on.The researchers found that only part of the characteristics of these features were recognizable and distinguish.The feature selection technique is to combine a number of low dimensional features and find the most effective feature vectors from it,which is beneficial to the final decision.In view of the above problems,this paper focuses on the research of feature extraction and feature fusion in content based image retrieval.As for feature extraction problem,we propose a novel feature extraction algorithm,CoCD(Contrast and Color Distribution),to represent color features in images.The original CoLD[1](Contrast and Luminance Distribution)features are mainly used to represent the color and texture features of the images.In order to obtain more effective information,we use the HSV color space instead of the original luminance distribution,and obtain better experimental results.In the aspect of feature selection,we use the LDA(linear discriminant analysis)to project the merged features into a supervised learning subspace,so as to get a more discriminative and lower dimensional feature descriptor.In this paper,Corel 5K(5000 images)is used as a test database,and the common Euclidean distance is used as a similarity measure.The accuracy rate(Precision),recall ratio(recall),P-R curve and average precision rate are used as evaluation criteria of retrieval image.We verify the performance of the proposed feature extraction algorithm and feature selection method is higher than other comparison methods.Experimental results show that the new feature extraction algorithm has improved significantly compared with some underlying feature algorithms,and the proposed feature selection method can also significantly speed up retrieval and improve retrieval efficiency.