Research on Fine-grained Proposed Region Extraction Method Based on Weighted Channel Network

Abstract

Proposed region extraction is an important step in target recognition and has important influence on subsequent results. While fine-grained images are more difficult to extract proposed regions due to the intra-class diversity and inter-class similarity. In order to solve the problem of fine-grained target proposed area extraction, This paper proposes a novel coarse-to-fine Weighted channel network(WCN)-based fine-grained image suggestion region extraction method, which firstly initializes parameters on the coarse-grained big data set, then the fine-grained data set is fine-tuned for specific problems to reduce model dependence on large-scale coarse-grained images, and finally the invalid features are suppressed while improving the effective features according to the response graph of the extracted features and the correlation of the feature channels to get the proposed region. The model was validated in the publicly available fine-grained image library CUB200_2011 and Stanford Dog, and achieved an accuracy of 80.1% and 82.8%, respectively, which just proves the validity and accuracy of the model.