Image Saliency Analysis Based On Multi-Instance Learning

With the rapid development of the internet, the online images are growing at an amazing rate, and the available computing resources are very limited with respect to such a mass of visual information. Therefore, people are eager to find a selection model which accords with human visual system to distribute limited resources to important information. According to this requirement, saliency detection on images based on selective attention mechanism has gradually become a hot issue in the field of computer vision.In this paper, multi-instance learning is proposed to analyze image saliency. The main work is as follows:1. A single-image saliency analysis algorithm is proposed. In order to retain the characteristic of the mutation region, relative features is used to describe the instance. Combined with EC-SVM, a classification model can be obtained by learning from labeled images. Comparing with the existing approaches, the proposed method achieves a better performance in saliency detection.2. A co-saliency analysis algorithm is proposed. A negative bag selection strategy is used to avoid the error label of the background area. In order to ensure the salient region prominent or noticeable with respect to its surroundings, an improved diverse density multiple-instance learning algorithm is proposed. The weights of instances are given by their single-image saliency map. The multi-image saliency map is generated according to the diverse density value of each instance. Experimental evaluation on the iCoseg image datasets demonstrates that the performance of the proposed algorithm is better.