Comparative Analysis of Fine-Tuned MobileNet Versions on Fish Disease Detection

Smart aquaculture farming is gradually becoming more popular and widely adopted in many fish farms, ranging from small to medium scale. To ensure that the final harvest quantity is not significantly reduced due to diseases, early detection of diseased fish individuals is crucial to remove them from the farming environment, thus preventing rapid spread to other individuals. The main objective of addressing this issue is to utilize identification and classification technologies as alternatives to manual classification methods, which help save time and significantly reduce costs in aquaculture. The various versions of MobileNet are leveraged to classify fish disease. MobileNet is a lightweight network due to its architecture, which can be analyzed depthwise, reducing the model size and computational cost. Training and evaluation were conducted on the SalmonScan dataset; the input image was resized to \(224\times 224\) , which several research teams had previously used. Experimental results show that MobileNetV1 achieves effective disease classification on fish with (96.30%) and (95.47%) accuracies, respectively, with and without data augmentation, indicating success in identifying disease symptoms in fish. Moreover, compared to previous studies on the dataset where manual processing was required to extract features from images to clarify disease points on fish, MobileNets demonstrate more robust support as it can autonomously learn these features on the network without the need for such complex processing.