Scene Classification And Objects Detection Based On Multi-color Space And Statistical Histogram

With the rapid development of computer and communication technology, multimedia technology is also changing quickly; the content of entertainment on network is mainly from word and picture to video. Network can supply to people the rich and colorful platform of video program, but it also is convenient for the objectionable videos propagation. Now, building a harmonious and open platform for the youth becomes the focus problem. At present, objectionable information detection technology can filter the network address, picture, word, and so on. The detection of video and audio is not yet perfect. Detection of objectionable videos is a challenging task, including the knowledge of multi-disciplinary and multi-fields. Therefore, how to detect objectionable videos efficiently and quickly becomes an urgent problem.In this paper, typical scene classification and object detection are the basic work in objectionable videos detection. Objectionable videos usually occurs in a particular scene, it is usually formed by the perspective of different objects or different objects. Scene classification is beneficial to understand video content and pertinence analysis the video content. Accurate classification can easily determine the scene where the event occurred and guide to adjust the sensitivity of the video. Especially the interior scenes need to pay attention to. Detection of object in and out can help to analyze the information correlation among the shots in the same scene. Currently, Our research group have made good results in shots segmentation and video style classification. Shots segmentation and scene segmentation are the basis of video analysis, the accuracy of shot segmentation will directly affect the accuracy of the typical classification. Video style classification determines the overall color style, and convenient to pertinence adjust the skin color models.In this paper we focus on resolving several basal issues of objectionable videos detection, the main research contents are as follows:1. Perfecting the video comprehensive analysis platform video based on multiple color spaces. The platform can display the video, choose different color space components, real-time display and calculate each frame single frame histogram, difference histogram and average scene histogram. Average scene histogram is mainly used for scene classification, scene classification module can extract the peak parameters features and achieve scene classification. Difference histogram is mainly used for object detection, object detection module can count the adjacent frame, by setting the threshold can achieve object detection. The platform can also be used to detect shots switching, video color style classification and the selection of effective color components.2.We have realized the classification of typical video scene by using muli-color comprehensive analysis platform. Typical scene often contain multiple shots, and these shots often cover all aspects of the scene; so we propose a new histogram which is formed by calculating the sum total of a color histogram of all the frames image of the video, the new histogram has good stability and could reflect the unique nature of the typical scene. But the histogram of different scenes often are different. In order to apply easily, we calculate the mean of the cumulative histogram, also named average scene histogram. The histogram describe the scene easily and effectively. In this paper we improve on extraction method of multi-peak histogram parameters, and achieve outdoor scene classification and describe the style of indoor scene by using the relevant classification rules. The result is good.3. We have realized object detection by using muti-color comprehensive analysis platform and frame difference histogram. The scene in the video often change slowly, the object always changes. Reflected on the histogram, the histogram of two adjacent frames changes little when there is no object in the scene, otherwise, it will significant change. Use the relationship of the histogram, we realize detecting the object in or out and determining the number of objects in the uniform background or small changes in the background. Currently, we have done only basic research, and detection results are not stable. In the following work, we will thoroughly analyze the regularity of histogram of frame difference to improve detection precision.