Multi-scale RA-CNN for fine-grained Image Classification

Fine-grained category identification is challenging and an inevitable subject of machine learning due to the problems of separating region localization and fine-grained feature learning. The necessity to reconcile accuracy standards with model optimization is inescapable in this area. Existing techniques mostly handle these issues separately, neglecting the fact that area recognition and fine-grained feature learning are interconnected and can so reinforce one another. The origin model introduced a new recursive attention convolutional neural network (RA-CNN) that recursively learns to differentiate regional attention and region-based feature representation at several scales and reinforces each other. The original network focuses on the transitivity between layer and layer, and it gives fine-grained identity strong guidance and good results. Our model doesn’t give up the global features and attention to extract the features in the process of the premise, can get better effect. From the point of experimental data, Multi-Scale RA-CNN combines the features of the three scales mentioned in the original RA-CNN, will bring higher accuracy.