Fish species recognition using an improved AlexNet model

The ability to differentiate between various species of fish plays an essential role in protecting its population, as well as the fishery industry. However, traditional machine learning algorithms are incapable of recognizing images with complex backgrounds and various illuminations. These fish algorithms are based on factors of color, texture, and feature extraction, which results in low recognition accuracy or inferior robustness. This paper aims to improve the fish recognition algorithm based on AlexNet by using the following methods: Firstly, item-based soft attention mechanism is utilized to facilitate accuracy. Secondly, the proposed model has less structural complexity, as it only consists of four convolutional layers, one item-based soft attention layer, and two fully-connected layers. Transfer learning is implemented to achieve a shorter training time. The experimental results show that our proposed model achieves higher accuracy and less computational complexity compared to many state-of-the-art fish recognition algorithms.