Research And Implementation Of Multi-feature-based Image Retrieval Technology

With the development and improvement of online shopping, this system has become more and more popular in consumers on their daily lives. In the online shopping system, consumers choose what they want by looking at photos of products'samples. Under this circumstance, lager quantity needs of image retrieval emerge. Every commodity will be taken at least one picture separately by the business, and what we care is the clothing area either. For what is mentioned above, we propose a method of image background elimination which is based on edge detection, in order to avoid the influence of background during the image feature extraction. The business may rotate the clothing when photo them so as to make the photos better looking. Consequently, the algorithm of image retrieval that we proposed must be robust for the rotating. According to the features of the clothing images, the shape feature of a clothing image as its attribute could be used to decide which class this clothing belongs to. The color and pattern of clothing are distinguished by the color and texture feature. So using a single feature could not describe a clothing image generally. And an image retrieval system based on single feature can not meat the requirements of the users. Then an image retrieval algorithm combined with color, texture and shape feature has been investigated in this paper.The methods of feature extraction and similarity measure of clothing images have been investigated as the main line. The domestic and foreign content-based image retrieval algorithms on clothing image features have been discussed in detail. Based on it, the description and extraction method of the visual characteristics of the descriptor, as well as measure of similarity, have been studied in priority. The clothing images that we chosen are the photographs of tiled cloth. In view of the clothing image features of their own, the invariant moment and the Fourier descriptor are proposed to retrieve the shape features of image. The HSI color space, which is closer to human vision characteristic, is quantized to become 72-dimensional color quantities. This color histogram is used to describe the color feature of images. The LBP operators are modified to be robust for rotating, then they can be used to describe texture feature of images. In this algorithm, the image retrieval based on shape feature is carried out at the first step, then the image background elimination. At last, the image retrieval based on color and texture features is accomplished using the image database that obtained at the first step.Moreover, a trademark image retrieval prototype system based on the multi-features has been constructed. The system can achieve the automatic feature extraction, similarity matching and the quadratic retrieval on the basis of color, texture and shape feature in the clothing images using VC++6.0. The experimental results demonstrated that the image retrieval algorithm based on multi-features is more practical and effective for applications than the algorithm based on a single feature. The multi-features algorithm is compared with the algorithm with single feature. It is shown that the retrieval speed of the proposed algorithm using the background elimination is faster than that of the algorithm without the background elimination algorithm in the condition of same image database and same similarity measure method. The image retrieval method with background elimination can both increase the speed of image retrieval and improve the quality of the system. The image retrieval algorithm based on multi-features is feasible and effective for clothing image retrieval. The research and practice in this paper have a certain reference value for the development and the promotion of clothing image retrieval based on multi-features.