Abstract:

Focusing on the problem of natural images scene semantics retrieval, a novel method based on multi-instance learning (MIL) is proposed. In order to transform the image retrieval problem into an MIL problem, first, an adaptive JSEG image segmentation method is designed according to the color complexity of images, and each image is segmented into several different regions, then each image is regarded as a multi-instances bag, and the color-texture features of each segmented region is regarded as an instance in the bag. Finally, an improved earth mover distance is used to measure the overall similarity among multi-instance bags (images), and a new lazy MIL algorithm for scene image retrieval is proposed. Experimental results on the COREL dataset show that this algorithm is feasible and the performance is superior to other algorithms.