Abstract: In this work, propose a system that can be sorted easily with fine-grained images. Do not use any comment objects / parts (weak supervisors) during training or testing, but only the training images of class labels. Sorting fine-grained target images for classification, only subtle differences between objects (e.g. dog's two breeds that look). Most of the existing works rely on the detector items / components to construct the correspondence between the parts of the object, At the very least, you need to train the exact annotation object or part of the object in the image. It points out the need for expensive items to prevent the widespread use of these methods. Instead, we propose to generate useful parts for the proposed size of the proposed component from the object, choose the recommended useful parts and the overall image representation for the calculation of the classification. This is specifically designed for classification fine-grained supervision because it has proven useful to play a key role in dependent works, but the exact detector is difficult to obtain part of the task. With Delegate suggestions can detect and visualize key parts (more discernible) Different types of object images. In the experiment, the proposed method of weak supervision to achieve a considerable or relatively weak advanced control method is more accurate, most of the existing methods Comments on the dependency of the three groups of challenging data. Its success shows that it is not always necessary to learn the expensive detector items / parts fine-grained image classification.

Keywords: classification, fine-grained, image, widespread, visualize key.