Deep learning-based material classification for waste management using acoustic spectrograms

This paper presents a robust deep learning framework for efficient material classification in waste management, focusing on a comparative analysis of acoustic time–frequency representations. Six commonly discarded materials (plastic, cardboard, steel, porcelain, glassware, and aluminum) were characterized through carefully preprocessed recordings. Signals were transformed using four techniques: Short-Time Fourier Transform (STFT), Continuous Wavelet Transform (CWT), Mel spectrogram, and Constant-Q Transform (CQT). A custom-designed, lightweight Convolutional Neural Network (CNN) achieved a peak classification accuracy of 97% with Mel spectrograms, significantly outperforming established pretrained models including VGG19 and MobileNet. To ensure methodological rigor, the framework was subjected to 5-fold cross-validation and statistical significance testing. Furthermore, a noise robustness analysis demonstrated that while Mel spectrograms excel in low-noise settings, the CQT representation provides superior resilience under severe industrial noise (0 dB SNR). External generalisability was validated using the public ESC-50 dataset, where a few-shot edge calibration protocol successfully rescued the model from an initial domain shift to achieve 77.50% accuracy. These results establish a hardware-aware, acoustic-only foundation for real-time automated recycling systems, demonstrating high adaptability for deployment on resource-constrained edge hardware.