Salient Region Detection And Its Application In Clothing Retrieval

Salient region detection is refer to that under the certain indicators or definitions to automatically detect the target that deemed to be salient. In recent years, it has been one of the very important research in computer vision field.According to its detection mechanisms, salient region detection can be divided into two categories: one category is bottom-up mechanism, based on visual attention; the other is mechanism of top-down. Most existing methods belong to the former, mainly rely on mutual contrast of visual features and due to lack of priori cognitive information makes this kind of methods always generate lower detection result, meanwhile to a particular task it can’t produce efficient detection results. Therefore, proposed a task-based high effective salient region detection method and at the same time considering human cognitive processes, it’s an urgent problem.In this paper, we simply combined salient region detection and target detection/ recognition, to propose a significant region detection method for clothing images, at the same time, we collect and established a clothing image salient region detection reference library to verify our method. The laboratory results expresses that our method can obtain higher accuracy. Besides, we have applied it to content-based clothing image retrieval system. The paper’s main work is as follows:1 After comprehensive consideration of varieties factors, we collect more than two thousand pieces of clothing images from major shopping website to build our own clothing image salient region detection reference library to validate the proposed approach and meanwhile provide benchmark to subsequent similar methods.2 On the basis of this reference database, five different universal salient region detection methods were compared, and analysis of their advantages and disadvantages.3 Taking into account the characteristics of the clothing image itself, face detection was used,then based on this face area to detect the area of clothing.4 Also according to the characteristics of the garment image itself, use HOG features, meanwhile using support vector machine(SVM) to effectively learning and training, then obtain detector to get salient region of clothing image.5 In order to improve the final accuracy of clothing detection, combine three different detection methods, namely serial combination of general salient region detection and object recognition/detection to obtained satisfactory results.