An Optuna-Driven Evaluation of Convolutional Neural Network Architectures for Biodegradable Fabric Classification toward Edge Deployment

The increasing volume of textile waste has intensified the need for intelligent sorting systems that can support recycling and circular-economy applications. This study presents an Optuna-driven comparative evaluation of five pre-trained convolutional neural network (CNN) architectures for biodegradable fabric classification toward edge deployment. The Biodegradable Fabrics Dataset contained 900 images from six classes, namely Abaca, Cotton, Hessian, Linen, Silk, and Wool, with 150 images per class. Images were resized to 224 × 224 pixels and partitioned using a stratified 70:15:15 training, validation, and testing split. MobileNetV3-Small, ResNet50V2, DenseNet121, EfficientNetV2B0, and InceptionV3 were evaluated under a common transfer-learning pipeline with frozen ImageNet pre-trained backbones. Optuna was used to optimize learning rate, dropout rate, batch size, and optimizer type using eight trials per architecture, with each trial trained for up to eight epochs and validation accuracy used as the objective function. The optimized models were converted to TensorFlow Lite (.tflite) format and evaluated using the TensorFlow Lite interpreter in the experimental environment based on post-conversion accuracy, model size, and inference time. ReseNet50V2 achieved the highest test accuracy of 94.81%, followed by InceptionV3 at 92.59% and MobileNetV3-Small at 91.11%. However, MobileNetV3-Small provided the most favorable deployment-oriented trade-off, producing the smallest size of 3.57 MB and the fastest inference time of 45.27 ms/image. All architectures preserved nearly identical accuracy after TensorFlow Lite conversion. These findings indicate that ResNet50V2 is the strongest accuracy-oriented model, while MobileNetV3-Small is the most practical candidate for edge-oriented biodegradable fabric classification.