<![CDATA[Automatic marine fish species identification plays an important role in supporting fisheries resource management, catch monitoring, and improving the efficiency of the identification process. The development of Convolutional Neural Networks (CNNs) has enabled automated image analysis with high accuracy. However, most CNN architectures, such as ResNet and Inception, have a large number of parameters and high computational requirements, making them less suitable for deployment on resource constrained devices. MobileNetV3 was developed as a lightweight CNN architecture, yet previous studies comparing MobileNetV3 Small and MobileNetV3 Large have reported inconsistent results, particularly for Indonesian marine fish species classification. This study aims to compare the performance of both MobileNetV3 variants for Indonesian marine fish species classification. The results show that MobileNetV3 Large achieved slightly higher accuracy, precision, recall, and F1-score than MobileNetV3 Small. However, the performance improvement of approximately 1% was not proportional to the substantial increase in the number of parameters, model size, and inference time. MobileNetV3 Small has 1,030,778 parameters, a model size of 5.18 MB, and an inference time of 2030.26 ms, while MobileNetV3 Large has 3,137,162 parameters, a model size of 13.7 MB, and an inference time of 2270.31 ms. Based on these findings, MobileNetV3 Small provides a better balance between classification performance and computational efficiency, making it more suitable for deployment on resource constrained devices.]]>
