Research On Region Graph Object Detection Under Gaussian Video Summarization

The development and popularization of digital technology and multimedia have derived diversified information forms. As the techniques to analyze and process image information, digital video analysis and digital image processing in which the video summarization and object detection technologies that have a wide application prospects in military, aerospace, biomedical, communication engineering, and other fields are included, have developed into commonly used tools in the scientific research and man-machine interface. Currently, the research of object detection techniques has made great progress and special summarization, video synergism analysis and so on are becoming the trend or direction of summarization development. However, it still can’t be regarded as perfection. In fact, it lacks a widely used model as a reference in the techniques of video summarization research. If different user requirements, actual application, etc. are taken into account, the performance of summarization will be widely divergent. Object detection is still facing many difficulties and challenges, such as same category difference, viewpoint changes, illumination difference and complex background which increase difficulties to detect objects accurately and rapidly.By means of analyzing the research status and algorithms about the existing video summarization and object detection field and summarizing the problems existing in current algorithms, this paper made in-depth discussion aiming at feature extraction, video segmentation, key frames selection in the process of generating video summarization and some object detection methods. On this basis, the paper put forward a few innovative methods.In the aspect of video summarization, according to the complexity of video visual content and interframe difference, this paper proposed a video frames similarity function (VFSF) which describes the difference of every set of three adjacent frames in time sequence to extract features. Tridiagonal matrix and extreme idea are added into VFSF, which not only decrease the computing capacity but also retain the most important image information. Inspired by the feature analysis results, the paper put forward the use of gaussian method for video segmentation, obtaining unequal sized shots and conquering the deficiency of content distortion in isometric segmentation. In the part of key frame extraction for video summarization, the gaussian method and split-merge method are proposed respectively. According to VFSF and video segmentation results, gaussian method verifies the gaussian distribution of feature values for every segmented shot and selects the frame with larger feature value as the key frame which can best represent this shot to generate video summarization. Through the two processes of split and merge, the split-merge method achieves classification and clustering for every shot, and determines the number of key frames for summarization automatically. The experimental results show that not only the computing amount and memory occupancy have been improved but also the key frames extracted can effectively summary video messages. Not affected by local information, split-merge method shows good detection effect that can successfully detect shot boundaries and identify segmentation frames especially for the shots with large content changes.In the aspect of object detection, the overview about image segmentation based on graph theory is given firstly. On the basis of theoretical analyses of such as minimum cut algorithm in video segmentation and improving the image segmentation method based on region, this paper proposed an intelligent object detection method based on region graph which can be summarized into two parts. First, the new image feature representation method which integrates color rarity and local boundary attributes and extracts image features in YIQ color space through a fuzzy function is put forward. Then we structure pixels eight-neighborhood relationship graph using obtained features, and add a user action to identify seed pixels, achieving the interactive object and background segmentation. Second, improving the object segmentation algorithm based on graphs, genetic algorithm is used to optimize the evaluation function regarded as a soft constraint under the hard constraint condition of user action. By means of searching the optimal binary image segmentation, we get the segmented objects and backgrounds of global optimization, guaranteeing the best balance between the color and boundary. Experiments show that the intelligent object detection method based on region graph about the detection for object and background can contain multiple isolated parts. Especially it is more obvious dealing with multiple object images. The object regions detected using our method both have good global and local connectivity. For images with strong correlation, it is able to detect the connected multiple objects and it can also identify isolated object regions locally for separated multiple object images, which gives consideration to both globality and locality.