Multi-scale confusion and filling mechanism for pressure footprint recognition.

Footprint data have large intra-class variances and small inter-class variances; therefore, the key to the footprint recognition problem is to mine and learn the discriminative local details in the footprint images. In this paper, an algorithm based on a multi-scale confusion and filling mechanism is proposed to address the problem of footprint recognition from the perspective of fine-grained image recognition. Firstly, the pressure footprint image is divided evenly into several sub-regions, and the score of each sub-region is calculated by a joint confidence function. Secondly, using the filling mechanism of the Region Filling Module, the region with the lowest score in the split image is filled with a higher one for data enhancement. Then, the filled image is confused once using the Multi-Scale Region Confusion Module, and the regions with high confidence score are confused again to obtain an image with multi-scale information. Finally, the footprint features of the filled image and the confused image are extracted by the backbone network and optimized by the joint loss function to carry out the task of footprint recognition. Comprehensive experiments show that the proposed algorithm achieves 89.3%, 93.4% and 86.5% on three benchmark dataset including CUB-200-2011, Aircraft and Stanford Dogs. Meanwhile, it obtains 97.8% on the footprint dataset. [ABSTRACT FROM AUTHOR]