Image Retrieval Based On Texture And Color Features

With the development of multimedia information, a large number of digital images have been created in the areas of entertainment, commerce, academia, and hospitals, etc; Image retrieval is becoming a hot research area. In the past, keyword indexing was the only way to search images; however, text-based image retrieval systems require manual annotation of images since it's not feasible to generate descriptive text automatically. Obviously annotating images manually will take a lot of labor for large-scale image database, and it's subjective and incomplete. But Content-Based Image Retrieval (CBIR) can solve this problem efficiently.CBIR is different from text-based image retrieval; it provides a way of content-based searching. CBIR uses the visual contents of an image such as color, texture, shape, etc to denote the image. In fact, it is a system which can retrieve images automatically by extracting some features from images and looking for the similar images which have the similar features. We compared color features, texture features and shape features, since texture contains important information about the structural arrangement of surfaces and the relationship to their surroundings, better suits the macrostructure and microstructure of the images. As a result, the paper dose research based on texture features, and tries to combine it with color features to retrieve images.At first, the paper starts with the research background and significance, introduces the current research situation and direction of CBIR, and then discusses the fundamental principle, frame, key technique, performance evaluation, relevance feedback technique and application fields.Then, the paper describes texture and texture features in detail; introduces some common methods to extract texture features, including grey level histogram moment, grey level co-occurrence matrix and Gabor wavelet. Since grey level co-occurrence matrix denotes the spatial relationship of grey levels of the texture in the images very well and its calculation speed is fast, the paper decides to extract the texture features by grey level co-occurrence matrix, normalize the feature vectors by Gaussian normalization method, and then retrieve the images by measuring the similarity. The result of experiments shows, the precision of this method is close to Gabor wavelet, but the computation complexity is low and the speed is fast, which means this method can meet user's needs.At last, in order to compensate the loss of color information, the paper introduces the methods to extract color features, color space and quantization methods. Since HSV space suits human understanding and memorial habit very well, so the paper quantifies the 3 channel H, S, V in accordance with human eyes'visual resolution based on the color quantization method, then extracts the color features by HSV histogram, and combines with the texture features extracted by grey level co-occurrence matrix to retrieve images. The result of experiments shows, this method can use color features and texture features efficiently to find those images which have similar color and texture features to the sample image.