Intelligent Television Video Analysis Based On Computer Vision Methods

With the development of science and technology, intelligent video analysis technology is developing fast. In recent decades, it has been widely used in intelligent city project, intelligent transportation, criminal investigation and handling etc. Video Analysis is based on the feature of the frame image. So firstly, we should extract appropriate feature to describe the image, and then analyze the feature by using the knowledge of image processing and machine learning. This thesis mainly focuses on the scene analysis of music video and some television video analysis such as image search and the detection of faces and logo. The main work of this thesis includes the following aspects:(1) We detect the abrupt and gradual edge of the video according to the similarity between the two adjacent frames which is measured combined the HSV histogram and surf features of the images. And then extract the key frames to be the content framework of the video. The experiment result shows that the key frames we get can effectively express the abstract of the video.(2) We use the classification algorithm retrieve some images of the video. First, we choose some images which is fulfill the retrieval characteristics to be train set data. Then let the frame images of video as a test set to do image classification. The experiment shows that we can get a good retrieval results by selecting an appropriate train data.(3) We obtain the neural network model of face detection by using a big face data set to train the AlexNet model. Then the frame images are detected using sliding window method to detect whether a face is included in these images. The experiments show that using the proposed face detection algorithm we can capture most of the video images containing human face.(4) At last we build the logo templates based on the edge and color firstly according to the characteristics of logo. And then identify the logo of the video through template matching method. We also discuss some related issues that affect the detected results and give the corresponding solutions. By experimental verification, the proposed logo detection algorithm in this thesis can get good results.