WCSE

Fine-grained Classification Using Multi-channel ResNet

Abstract— At present, fine-grained classification has attracted extensive attention. The task of fine-grained classification is difficult due to the challenge of accurately locating the key regions with resolution and extracting valid features from the detected key regions for classification. In this paper, we propose a new convolutional neural network (Multi-channel ResNet). Multi-channel ResNet uses Mask R-CNN for foreground extraction to reduce the interference of image background on fine-grained classification results. In addition, the four-channel ResNet module is used to learn fine-grained features at multiple scales, and Gaussian blur processing and crop processing are used to learn details and contours, all and local features, so as to improve the accuracy of fine-grained classification. The model does not require bounding box/part annotations. We experiment with the CUB_200_2011 dataset, and the results show that Multi-channel ResNet has an improvement in fine-grained classification tasks on the baseline of no pre-trained ResNet-18.

Index Terms— Fine-grained Classification, Multi-channel ResNet model, Convolutional Neural Networks

Di Zang, Yiqing Yan, Jun Chen, Yang Li
Department of Computer Science and Technology, Tongji University, CHINA
College of Civil Engineering, Tongji University, CHINA
State Key Laboratory of Disaster Reduction in Civil Engineering, Tongji University, CHINA

[Download]

Cite: Di Zang, Yiqing Yan, Jun Chen, Yang Li, " Fine-grained Classification Using Multi-channel ResNet " Proceedings of 2020 the 10th International Workshop on Computer Science and Engineering (WCSE 2020), pp. 37-41, Shanghai, China, 19-21 June, 2020.