Research on fine-grained pattern recognition based on attention pattern-generated model

3rd International Symposium on Resource Exploration and Environmental Science

生态环境科学

Zhou, Jiayin^1 ; Feng, Kaiping^2 ; Luo, Lihong^2

School of Computer Science, Guangdong University of Technology, Guangzhou Guangdong

510006, China^1

School of Art and Design, Guangdong University of Technology, Guangzhou Guangdong

510090, China^2

关键词: Activation volume;    Fine grained;    High-dimensional;    Multi scale prediction;    Pattern recognition algorithms;    Recognition accuracy;    Stanford;    Visual differences;   

Others  :  https://iopscience.iop.org/article/10.1088/1755-1315/300/2/022038/pdf
DOI  :  10.1088/1755-1315/300/2/022038

学科分类：环境科学（综合）

来源: IOP

【 摘 要 】

The major difficulty in classifying the type of fine-grained patterns is that the visual differences between subordinate types are probably very subtle, which, however, may be extremely big in the same kind. A fine-grained pattern recognition algorithm is put forward based on the attention pattern-generated model in this article. First, extract high dimensional characteristic pattern from the image via CNN. Then apply several separators to outputting category predictive activation volumes on the characteristic mapping. After that, polymerize these activation volumes to obtain attention mapping. As for the fine-grained recognition, we use the generated attention pattern to tailor and zoom the intentional area so as to carry out multi-scale prediction. This article has been assessed on CUB-200-2011, FGVC-Aircraft, Stanford Cars, and the model mentioned in this article is more superior than baseline multi-scale resnet50 by more than 2%. Attention pattern generated from the model is helpful for other models to increase their recognition accuracy as well.