Image Classification Based on Weight Adjustment before Feature Pooling

Shaokun Feng, Hongtao Lu, Lei Huang

Abstract: In image classification based on Bag-of-Features(BoF), the Locality-constrained Linear Coding (LLC) is a successful implementation, which is a more effective coding scheme compared with the traditional vector quantization(VQ) coding. Although, to achieve the best performance, max pooling scheme is chosen in the SPM layer, much of the spatial information is still lost during the pooling step, because all the coded descriptors are given the same importance to obtain the final representation. In this paper, we propose a new scheme that makes full use of spatial structure information to readjust their relative weights red and thus give some descriptors more chances to appear in the final feature vector more than others. Experiments of image classification on benchmark datasets show that the proposed method outperforms the LLC method.