Evaluasi Efisiensi Komputasi Model MobileNetV2 Berbasis TensorFlow Lite untuk Klasifikasi Kesegaran Tiga Spesies Ikan pada Perangkat Edge

Abstract

Fish freshness classification using computer vision offers a potential approach for rapid visual assessment. However, deploying deep learning models on edge devices requires consideration of both classification performance and computational efficiency. This study evaluates MobileNetV2 implemented with TensorFlow Lite for six-class image classification representing fresh and non-fresh conditions in three fish species: Johnius heterolepis, Rastrelliger brachysoma, and Nemipterus bathybius. The dataset described in this study consists of 372 images, comprising 300 training images and 72 validation images. Based on the confusion matrix presented, the model correctly classified 69 of the 72 validation images, yielding an accuracy of 95.83%. Misclassifications occurred in the non-fresh short mackerel class and the fresh large-scale croaker class. The computational evaluation framework considers model size, inference latency, throughput, and classification accuracy across Float32, Float16, and INT8 configurations. However, the efficiency values currently presented are illustrative scenarios rather than verified measurements obtained from converted models on a specified edge device. Therefore, no definitive conclusion can yet be drawn regarding the most computationally efficient configuration. Further experiments involving actual model conversion, consistent evaluation procedures, and device-level measurements are required. This study highlights the importance of evaluating classification performance alongside computational requirements when developing image-based fish freshness classification systems for edge deployment. The findings should be interpreted within the limitations of the available dataset and its labeling procedure.