MobileNetV2 Pruning for Improved Efficiency in Catfish Classification on Resource-Limited Devices

The emergence of new deep learning architectures, including ResNet, VGG, and AlexNet, has resulted in enhanced performance, increased network architecture complexity, and high computational requirements. In the context of deployment on devices with limited resources, such as mobile devices or edge computing, these deep-learning models are impractical. To address this challenge, this study proposes modifications to the architecture of MobileNetV2, a lightweight convolutional neural network (CNN) specifically designed for devices with limited computing power and memory. The modifications involved reducing the number of bottleneck layers in MobileNetV2 from 17 to 12, which aimed to reduce the complexity of the model without sacrificing accuracy. To maintain network efficiency, we also added additional layers like SeparableConv2D and GlobalMaxPooling2D. The test results demonstrate that this modification enhances the classification accuracy to 93.13% and reduces the number of parameters to 59.36% when compared to the baseline model. Additionally, the proposed model achieves an inference time of 186.83 ms, which is a 6.84% improvement over the original MobileNetV2’s inference time of 200.54 ms. This improves inference speed.