Fishnet: A Deep Neural Network for Fish Species Classification

Fish, a vital source of minerals and micronutrients, plays a crucial role in a healthy diet, especially in regions where the freshwater fish industry holds significant commercial value. Accurate fish classification is essential, as different species exhibit distinct signs of disease and decay. However, traditional identification methods are often inefficient. Deep learning offers a faster, more accurate solution. This study introduces FishNet, a deep learning model combining MobileNetV3, Self-Attention, and residual blocks, resulting in a lightweight architecture suitable for real-time applications. The model enhances MobileNetV3 by incorporating self-attention mechanisms to capture relevant spatial features and residual blocks to improve training stability and performance. Compared to existing models such as MobileNetV3Small, Xception, DenseNet121, InceptionV3, and ResNet50, FishNet achieves superior accuracy of 98.35% on the Fish Dataset (Cabuyao City) from Kaggle, while requiring fewer training images and computational resources, making it ideal for mobile and embedded devices. The goal of this work is to provide an efficient solution to prevent fraud in the global fish market by allowing accurate, on-the-spot fish classification.