Fine-grained image classification is to recognize hundreds of subcategoriesin each basic-level category. Existing methods employ discriminativelocalization to find the key distinctions among subcategories. However, theygenerally have two limitations: (1) Discriminative localization relies onregion proposal methods to hypothesize the locations of discriminative regions,which are time-consuming. (2) The training of discriminative localizationdepends on object or part annotations, which are heavily labor-consuming. It ishighly challenging to address the two key limitations simultaneously, andexisting methods only focus on one of them. Therefore, we propose a weaklysupervised discriminative localization approach (WSDL) for fast fine-grainedimage classification to address the two limitations at the same time, and itsmain advantages are: (1) n-pathway end-to-end discriminative localizationnetwork is designed to improve classification speed, which simultaneouslylocalizes multiple different discriminative regions for one image to boostclassification accuracy, and shares full-image convolutional features generatedby region proposal network to accelerate the process of generating regionproposals as well as reduce the computation of convolutional operation. (2)Multi-level attention guided localization learning is proposed to localizediscriminative regions with different focuses automatically, without usingobject and part annotations, avoiding the labor consumption. Different levelattentions focus on different characteristics of the image, which arecomplementary and boost the classification accuracy. Both are jointly employedto simultaneously improve classification speed and eliminate dependence onobject and part annotations. Compared with state-of-the-art methods on 2widely-used fine-grained image classification datasets, our WSDL approachachieves the best performance.
展开▼
