An image retrieval method based on feature fusion

With the exponential growth of image data, efficient and accurate retrieval methods are paramount. This paper introduces CVIH, a pioneering deep hashing algorithm that fuses CNN and an enhanced ViT to improve image retrieval. By extracting local features with a computationally efficient CNN and global features with a ViT improved by the Kansformer encoder, CVIH generates highly discriminative hash codes. A novel loss function addresses sample imbalance and hash code discreteness. Evaluations on CIFAR-10 and NUS-WIDE show superior performance compared to existing ViT-based methods.