Multi-audio feature maps fusion for watermelon quality detection

Abstract

Non-destructive detection of the internal quality of watermelons after harvest can significantly reduce losses and waste during the subsequent sales process. However, existing algorithms often struggle with limited generalization and high iteration costs. This study leverages audio feature maps of watermelons and employs deep learning to classify ripeness (ripe or raw) and internal defects (hollow or juicy). A hybrid attention mechanism, DWTR, is proposed to enhance feature extraction by adaptively capturing spatial and channel information. Additionally, re-parameterization branches are introduced to boost model representation without increasing inference overhead. The Rep-MBF model, a multi-branch fusion approach, was developed utilizing Mel spectrograms and short-time fourier transform (STFT) spectrograms of single-tap audio as dual inputs. The Rep-MBF model demonstrated strong performance with an accuracy of 97.81%, precision of 97.49%, recall of 97.35%, and F1-Score of 97.42% on the test set. The model's inference time on a Raspberry Pi 4B (8 GB) edge computing platform was only 16.24 ms, meeting the accuracy and speed demands for watermelon internal quality detection. In real-world detection scenarios, the Rep-MBF model accurately predicted 44 out of 48 watermelon samples, achieving an overall detection success rate of 91.67%, demonstrating excellent performance in practical watermelon detection applications. The Rep-MBF model achieves high-precision, low-latency detection of both watermelon ripeness and internal defects, while also demonstrating excellent robustness. These combined attributes provide strong algorithmic support for the development of portable nondestructive detection devices for watermelon internal quality.