Spectrogram-based Music Retrieval

As the wide spread of computers and the rapid development of Internet,digital music reaches the far corner of the world.This brings about the issue of how to find a song efficiently and effectively.Tag-based music retrieval is currently the only popular approach to access music.As a customer,you need to figure out the relevant tag information,while as a service provider, much labor is involved to manage the database.Meanwhile,troubles as follows usually haunt us:a clip of melody occurs to us but brings no identity information along with it.Therefore, the problems give birth to the content-based music retrieval(CBMR),which focuses on music features that can express its content,and thereby guarantees the retrieval accuracy.With an intensive research on spectrogram,we try depicting the music content implicitly and building a fuzzy "image" for each song based on it.A simple feature matrix is proposed to represent music texture which can amortize the impact of noise disturbance.Moreover,the feature matrix stores relative data instead of absolute data to enhance its robustness.We also propose a new similarity algorithm based on the music theory and the sound nature.It deduces the similarity degree of two features by dot-dividing them,setting weights and a normal interval for the elements and computing the rate of abnormal elements.A positioning algorithm of the anchor points will accelerate the speed of matching.MARBOS is built on the above theory and research.It takes query-by-example as input, portrays the music content in spectrogram feature matrix and deploys matrix similarity algorithm to match queries against candidates in the database.A group of experiments conducted to evaluate several characteristics indicates a better performance of MARBOS than one designed with the main stream idea.