Learning optimal image representations through noise injection for fine-grained search

In recent years, fine-grained image search has been an area of interest within the computer vision community. Many current works follow deep feature learning paradigms, which generally exploit the pre-trained convolutional layer's activations as representations and learn a low-dimensional embedding. This embedding is usually learned by defining loss functions based on local structure like triplet loss. However, triplet loss requires an expensive sampling strategy. In addition, softmax-based loss (when the problem is treated as a classification task) performs faster than triplet loss but suffers from early saturation. To this end, a novel approach is proposed to enhance fine-grained representation learning by incorporating noise injection in both input and features. At the input, input image is made noised and the goal is set to reduce the distance between the L2 normalized features of input image and its noisy version in the embedding space, relative to other instances. Concurrently, noise injection in the features acts as regularization, facilitating the acquisition of generalized features and mitigating model overfitting. The proposed approach is tested on three public datasets: Oxford flower-17, Cub-200-2011 and Cars-196, and achieves better retrieval results than other existing methods. In addition, we also tested our approach in the Zero-Shot setting and got favorable results compared to the prior methods on Cars-196 and Cub-200-2011.