Research On Visual Perception-Based Image Retrieval

Due to the fast and steady development of multimedia acquisition equipments, multimedia processing methods and the Internet, huge amounts of images/video data appeared, and were being shared, then large image databases began to emerge in many traditional and new-born fields. But, it is an important and challenging job to find a way to retrieve the user-wanted images from the image databases rapidly and effectively. Being a promising solution, content-based image retrieval (CBIR) technique received ever-increasing concern from almost every field, and has been a very active research domain.In this dissertation, all the key techniques of CBIR are briefly discussed, and the explorery research work focus on the human perception imitation. Because image retrieval is a substantially subjective recognition job, every link of CBIR should fully reflect the visual perceptual characteristics to increase the consistency between computer-imitation system and human visual system. Till then, the entire system desirable performance enhancement could be achieved. The researches are carried out in three aspects as follows: feature quantization, image feature acquisition and dissimilarity measures.The main contributions of the dissertation are summarized as follows:1. Firstly, an HSV color space quantization algorithm is proposed based on a gray-color boundary curve function.Because of the abnormity of color distribution in the HSV space, the common solution use predefined quantities for the space quantization; that leads to fairly quantized error especially in the low luminance and low saturation areas. But, in our method we avoid the quantization error by dividing the HSV space into gray zone and multi-color zone in light of visual perception practice. The method could be fulfilled in two steps: first step, the color values is converted from HSV space to CIE L~* a~*b~* space, and the curve points are located by obtaining the regional difference maximized ones; second step, the curve fitting is achieved by using least-squares method. Then a quantization algorithm is presented based on the gray-color boundary curve function. In addition, the curve function is basically independent with other quantization method, so it can work jointly with the others.2. Secondly, a color visual-attention function based feature acquisition...