Fish Species Identification on Low Resolution - A Study with Enhanced Super Resolution Generative Adversarial Network (ESRGAN), YOLO and VGG-16

Abstract

An intelligent detection and recognition model for the fish species from camera footage is urgently required as fishery contributes to a large portion of the world economy, and these kinds of advanced models can aid fishermen on a large scale. Such models incorporating a pick-and-place machine can be beneficial to sorting different fish species in bulk without human intervention, significantly reducing costs for large-scale fishing industries. Existing methods for detecting and recognizing fish species have many limitations, such as limited scalability, detection accuracy, failure to detect multiple species, degraded performance at a lower resolution, or pinpointing the exact location of the fish. Modifying the head of a compelling deep learning model, namely VGG-16, with pre-trained weights, can be used to detect both the species of the fish and find the exact location of the fish in an image by implementing a modified YOLO to incorporate the bounding box regression head. We have proposed using the ESRGAN algorithm and the proposed neural network to amplify the image resolution by a factor of 4. With this method, an overall detection accuracy of 96.5% has been obtained. The experiment has been conducted based on a total of 9460 images spread across 9 species. After further improving the model, a pick-and-place machine could be integrated to quickly sort the fish according to their species in different large-scale fish industries.