Video Commercial Recognition And Interactive Image Retrieval

With the development of multimedia technology, a large number of online videos are releasing and spreading. Commercial videos also have a variety of form and content. How to effectively manage the commercials in the massive videos is a problem to be solved urgently.In this paper, a novel framework based on support vector machine (SVM) for commercial automatic recognition and commercial segmentation is proposed to realize the supervision of advertising. The main work is as follows:1. On the basis of shot boundary detection, a new kind of context-based feature is proposed to build the temporal-spatial relations of shots by combing neighbor shots;2. We propose a novel SVM-based commercial recognition algorithm and a new rule-based post-processing method;3. We introduce a commercial segmentation method based on short-time energy audio feature to segment the commercial recognition results into slices.Experiments show that the recall of commercial recognition is above95%and the precision of commercial segmentation is more than80%.With the explosive growth of image data, content-based image retrieval (CBIR) has gained more and more attention in recent years, while the semantic gap and curse of dimensionality are still two open questions of CBIR. In this paper,we propose a new interactive image retrieval method based on locality-sensitive hashing (LSH) and support vector machine (SVM). The main work is as follows:1. We propose a new image retrieval method based on LSH. LSH is adopted as an index structure to overcome the curse of dimensionality and get initial retrieval results;2. We present a SVM-based relevance feedback (RF) scheme. RF is introduced to shorten the semantic gap between low-level features and high-level perception. In the RF process, we present a method to expand training samples and retrieval results to improve the accuracy of image searching.In order to demonstrate the effectiveness of the proposed system, we test our method on two image sets. The comparisons with existed methods show that the proposed method is satisfactory in both initial retrieval results and relevance feedback results, the precision of image retrieval will get95%after twice or three times feedback.