Robust and efficient recognition of low-quality images by cascaded recognizers with massive local features

For image recognition with camera phones, defocus and motion blur cause a serious drop of the image recognition rate. In this paper, we employ generative learning, i.e., generating blurred images and learning based on massive local features extracted from them, for a recognition method using approximate nearest neighbor search of local features. Major problems of generative learning are long processing time and a large amount of memory required for nearest neighbor search. The problems become serious when we use a large-scale database. In the proposed method, they are solved by cascaded recognizers and scalar quantization. From experimental results with up to one million images, we have confirmed that the proposed method improves the recognition rate, and cuts the processing time as compared to a method without generative learning.