Exploiting Best Practice of Deep CNNs Features for National Costume Image Retrieval

摘要

Convolutional neural networks (CNNs) have recently achieved remarkable success with superior performances in computer vision applications In most CNN-based image retrieval methods, deep CNNs features are verified as discriminative descriptors for effective image representation. This paper exploits the best practice for CNNs application to national costume image retrieval. Several important aspects that affect the discriminative ability of deep CNNs features are investigated thoroughly, including layers selection, aggregation and weighting methods. Firstly, an effective weighting method for sum-pooling features aggregation is given, which is more suitable for national costume image than some typical aggregation methods such as SPoC and SCDA. Secondly, in view of the complementary strengtis, compact multi-layer CNN features combined with low dimensions are proposed and proven to be effective for national costume expression. Finally, a re-ranking strategy of diffusion process is applied to further enhance the performance for national costume images retrieval. The experimental results show that the proposed method outperforms the existing methods remarkably, which will provide some new research ideas and technical references for researchers in the field of national costume image retrieval.