Research On Some Technologies Of Content-based Image Retrieval And Classification

With the development of technology and the rapid expansion of the Internet, there has been an increasing number of image data, followed by the emergence of a growing number of large image database, the network also carries vast amounts of image data resources.Since the 70's of last century, there have been text-based image retrieval. Text-based image retrieval developed from the text retrieval technology using text label of image.It's key technology is text retrieval. Until the 90's of last century, people began to study content-based image retrieval. Content-based image retrieval extract features such as color, texture, shape, spatial location of image automatically. Content-based image retrieval has obvious advantages: no longer labeling images manually, which means search results will not be affected by the subjective dimension; improve the efficiency of image retrieval.There are three kinds of image retrieval methods: content-based, semantic-based and feedback-based retrieval, and content-based image retrieval is the most popular one of all. Among various physical characteristics of the image, Color feature is one of the most basic features, and is widely used by retrieval algorithms. The color feature will not change when position or rotation situation of image change. However, as the color image is usually colorful, the dimension of color feature vectors is very high which resulting in increased computation. The time consumption become very large When the image set grows large. There are usually two way to solute this question. One is to define several color region, similar color is classified into same region.And the dimension of vectors will decrease accordingly. Another way is to calculate the histogram, choose the most frequent colors (dominant color), as feature vector.Image Classification is closely contacting with image retrieval. The key procedures of image classification include: image preprocessing(Image Enhancement, image denoising, and image restoration,etc.), feature vector extraction, build classification model and image classification. Three popular classifiers are: statistics-based, rule-based and artificial neural network based classifier. Support Vector Machine as a machine learning method developed from stastics method has great advantages in small sample, nonlinear or high-dimensional pattern recognition problems.This paper introduces three kinds of retrieval methods, discusses popular feature of image and the similarity measurement methods, explains two algorithms in detail: one is Image Retrieval algorithm, elaborates image retrieval algorithm integrating dominant color and color quantization, the other is Image Classification altorithm based on Support Vector Machine.The contribution of this paper is summarized as follows:1) This paper presents image retrieval algorithm integrating dominant color and color quantization. First, user of retrieval system should select one image as demo image, and also specify the region of interest(ROI) with rectangle. Second, we calculate histogram of ROI and choose dominant colors, then quantize color of demo image. Third, calculate the weight of every color region according to dominant color information, and finally form feature vector of demo image.2) This paper presents image classification algorithm. The algorithm integrates existing image segmentation method based on threshold and Support Vector Machine, presents rules for foreground image judging. Execution processs are: first, convert color image to gray image, segment demo image into several areas and find out the foreground image with judging rules, calculate vector of color feature of foreground image, at last train the Support Vector Machine. Conducts the research on deep Web resource search, designs a deep Web vertical search method.