This class project explores the implementation of the paper ”Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-grained Image Recognition” [1] on the Plant Pathology 2021’s dataset, a competition from Kaggle platform. The RA-CNN model will crop and up-sample the attended region in a coarse image to allow the model to recognize more features in details. The training strategy is used to iteratively and alternatively optimize weights in the classifier and APN (Attention Proposal Network) with pre-trained EfficientNet B0 CNN features extractor. An EfficientNet B0 has been used for feature extraction instead of a VGG-19, making training less memory greedy and less computationally expensive.
The paper of the project can be found at link to pdf .