MS-ALN: Multiscale Attention Learning Network for Pest Recognition

Complex backgrounds, occlusions, and non-uniform classes present great challenges to pest recognition in practical applications. In this paper, we propose a multiscale attention learning network address these problems. This recursively locates discriminative regions learns region-based feature representation four branches. Three newly designed modules, which are target localization, detection, removal connect two extracting sub-networks adjacent branches generate images of different scales. The localization detection modules locate the filter out complex backgrounds while module randomly removes region encourage model tackle occlusions. Thereafter, parameter-shared classification sub-network follows every branch for recognition. A decoupled strategy is adopted problem classes. We experimented on widely used IP-102 dataset achieved state-of-the-art performance.