Efficient sweet potato quality detection using pruned attention-based deep learning

Abstract

Sweet potatoes are rich in essential nutrients crucial for maintaining human health, and their market value is primarily determined by external quality. To address the classification of sweet potato appearance quality, we propose a rapid assessment method for evaluating external quality. Using a dataset of 1833 images of the Yanshu 25 variety, we adopt the MobileNetV4 series network with a reduced Universal Inverted Bottleneck (UIB) module as the core architecture. The network is trained and pruned by integrating a learning rate scheduler, a multi-head self-attention (MHSA) mechanism, and a pruning algorithm based on gated mechanisms. After pruning, the optimal model achieves an accuracy of 93.08 % across five categories, with an average detection speed of only 0.0022 s per sample. This method offers outstanding detection accuracy, faster processing speed, and significantly reduced model complexity, making it highly effective for sweet potato quality assessment.