Study On Key Techniques Of Content-Based Audio Retrieval (CBAR)

With growing development of multimedia and network technique, information is becoming rich day-by-day, and information retrieval technique is under grand change. People are not contented with traditional text-based retrieval, instead, they need a retrieval engine which supports the fast retrieval of multimedia data, such as video, image, and audio. Content-based audio retrieval (CBAR) is presented then for audios, which extracts semantic clues from media and retrieval media based on these semantic clues. In this way, the retrieval process is directly related with the semantic of media, and the flexibility and efficiency of retrieval is enhanced.In this paper, firstly, the developing process of CBAR is introduced, secondly, secondly, some key technology for a successful CBAR is described and an improved audio segmentation algorithm is presented, and then the results of retrieval experiments and analysis are given. Lastly, shortcomings and the direct of study in the future are put forward.Effective audio segmentation and classification are the preconditions of CBAR. The traditional threshold-based methods mostly adopt relatively simple feature and experienced values. The classification is single and the feature threshold selection is very difficult. Therefore, we adopt the improved Gaussian-based segmentation algorithm and present a new feature named Mel-ICA to improve it. This method does not need samples. The segmentation is implemented according to feature change-points, and good results are achieved. Besides, a classification approach based on threshold and models is proposed, which combines the advantages of these two methods. And the wave and Fourier transform are used to extract features, with the accuracy of classification increased greatly.Feature description is the key to the system. This study integrates the time domain, frequency domain and time-frequency domain analysis methods, to depict the essential of audio signal, and constitute the description operator. Query-by-example (QBE) is adopted for audio retrieval. Firstly, the (MST) is used for clustering to form key-frames, and the frames with same type are compared to get the similarity. As a result, the calculation complexity is decreased and the retrieval efficiency is increased greatly. The experiments can be concluded that our method can retrieve at the level of objects and achieve good performance. Finally, the paper summarizes the study and declares the direction of further research and exploration.