Salient Region Conducted Local Feature Algorithm

With the rapid increase of multimedia data, one of the most significant challenges is to understand and interpret such a huge amount of multimedia data. Currently, more and more retrieval applications are emerging to process these multimedia data, such as video recommendation, travel guidance and content-based TV copy identification. Among these systems, a fundamental step is to extract feature information from images. Image feature extraction algorithms include two main domains:global feature-based and local feature-based. While global features considering the feature of the whole image like color and texture, local features focus on the features in each subregion of an image, such as edges and corners. Although global feature-based algorithms can achieve a high processing speed, their accuracy is not guaranteed. On the other hand, compared to global feature-based algorithms, local feature descriptors are more robust, both scale-invariant and rotation-invariant, but need more computation and storage space for hundreds of feature vectors in one image. As a result, even the SURF (Speeded Up Robust Features), state-ot-the-art local feature algorithm optimized for efficiency, shows a very slow processing speed. While evaluated on a server with3.3GHz Core i7CPU, it can only achieve the processing speed of2.6frames per second, far from the real-time requirement. Furthermore, high-dimension vectors used to describe local features require large storage space, for images with hundreds of or thousands of local features need several hundred KB or several MB to store.This paper presents a novel method to reduce the computation of local feature algorithms with an optimized algorithm called SRLF (Salient Region conducted Local Featrue). This algorithm focuses on the salient region of images, in which the local features locate closer to each other and show more importance to the accuracy than others outside. In addition, the precise salient region boundary hurts the accuracy when used to filter local features, since most of local features just locate on the boundary of salient regions which is actually detected on objects’edges and corners. Thus, SRLF employs the geometric information of local features to locate the approximate salient regions and performs the high-dimension vector computation only on those small regions. This helps to reduce the total amount of computation with fewer local features. This salient region is computed from local features’ distribution. It first tries to divide the whole image into several blocks, where each block is a dense region of local features. Then the algorithm calculates the geometry center on each image block as the salient region centres. At last, with knowledge of the distribution shape of local features, the algorithm grows the region until a predefined threshold. In the evaluation, SRLF helps to reduce more than half of local feature in the original algorithm, which speeds up the whole algorithm with1.6X and reduces half of the storage space. Furthermore, after integrated to an existed image retrieval system, SRLF helps to speed up the whole system about2X.SRLF presents three main advantages when compared to other techniques:1) based on local features, SRLF can be integrated with local features to boost the whole image retrieval system.2) SRLF introduces no additional image features and computes very efficiently.3) As an approximate algorithm, SRLF shows no obvious precision loss in a realistic image retrieval system.