Waste Image Classification Using Fine-Tuned MobileNetV2 under Imbalanced Data Conditions

Abstrak

Ineffective waste management is a critical global environmental issue where manual sorting is often inefficient and prone to human error. Deep Learning technology, specifically Convolutional Neural Networks (CNN), offers automated solutions to improve classification performance. This study aims to optimize waste image classification by evaluating the effectiveness of the Fine-Tuned MobileNetV2 architecture, specifically addressing the challenges of imbalanced data distribution where recyclable items significantly outnumber residuals. Experiments compared six training scenarios including Basic CNN, CNN with Class Weight, and Transfer Learning using VGG16 and MobileNetV2 with frozen and fine-tuning strategies. Evaluation metrics included Accuracy, Precision, Recall, F1-Score, and Confusion Matrix analysis. Results indicated that the Basic CNN model struggled with the minority class, yielding a low F1-Score of 0.49. Conversely, the optimized MobileNetV2 Fine-Tuned model achieved superior performance, recording a testing accuracy of 96.49% and an F1-Score of 0.92. It is concluded that the fine-tuning strategy on MobileNetV2 is the most effective approach to mitigate data imbalance, providing an optimal balance between high sensitivity toward the minority class and computational efficiency for real-time implementation.

Identifier

• DOI: https://doi.org/10.30871/jaic.v10i2.12218
• OpenAlex: https://openalex.org/W7160864104