Fast Local Feature-Based Saliency Region Detection Algorithm

With the rapid popularization of Internet and cloud computing, multimedia data, such as video and image, has become one of the major data types need to be processed. Unfortunately, many multimedia applications, such as image search engines and network information filtering systems are far from applicable, mainly because the multimedia retrieval algorithms which these applications are based on need long time to process mass of information included in multimedia. Image features extraction is important to multimedia retrieval problem. However, extracting features from an image generates huge amount of information, which limits the processing time. As human only focus on the saliency region of the image, detecting non-salient region is unnecessary. Therefore, detecting saliency region before extracting image features which excluding non-salient region is a good choice to reduce feature information and improve processing speed. The traditional saliency region detection algorithms are too complicated, resulting in large overhead. This paper introduces a fast salient-region detection algorithm which combines feature points distribution and max sum of sub-matrix algorithm. This algorithm can detect saliency region efficiently and lose less precision, which can be applied for image retrieval system. The contributions of this paper are:●Utilize local feature points distribution which generated on feature detection stage to detect saliency region efficient and lose less precision(?) Utilize the coordinate of local feature points to generate the distribution matrix of feature points. The distribution matrix describes the distribution of feature points.(?) Utilize the standard deviation of distribution matrix to indicate discrete degree of feature points.(?) Define a non-negative adaptability matrix which is related with the number of local feature points and the discrete degree of feature point distribution(?) Utilize the idea of dynamic programming to compute the sum of max sub-matrix efficiently, which can detect saliency region of image efficiently.(?) According to the statistics, weights matrix can be used to improve the precision of algorithm.(?) This algorithm can be expand to compute N-max sub-matrix, which means N saliency region can be found.●As the feature information has been reduced after detecting saliency region, the speed of matching stage of multimedia retrieval algorithm would be improved.