Abstract
<title>Abstract</title> <p>Plastic waste pollution continues to be an ecological problem, making it essential to have precise and effective automated waste sorting solutions. Even though image classification models based on deep learning demonstrated promising results, the scarcity of domain-related benchmark datasets hinders model generalization and classification. This research contributes to the creation of a PLASTIC dataset which includes 4,500 images in 15 classes of plastic waste. Ten of the most popular architectures of convolutional neural networks (CNN), such as ResNet, VGG, GoogLeNet, EfficientNet, and AlexNet, are compared by means of a 70:20:10 train/validation/test split. Comparative experiments conducted on a general-purpose dataset of garbage prove much worse results, illustrating the importance of a domain-related plastic waste dataset. Moreover, this research proposes Adaptive Training Controller (ATC), which is a generalized, transparent, and architecture-independent autotuning algorithm improving training via adaptive learning-rate tuning, intelligent optimization, mixed precision computation, and hardware awareness resource management. In contrast to Optuna, ATC can be used with most of the CNN architectures without any architecture-specific adjustments, thus minimizing the number of GPUs required for training, decreasing time and memory requirements, and improving stability. The experimental results illustrate that ATC always yields the highest accuracy values compared to Optuna with 94.89% using ResNet101 and 94.22% using ResNet50. From the results, it can be seen that ATC is a fast and effective autotuning solution for deep learning-based plastic waste classification.</p>