GarlicNet: A pressure signal-based CNN-Transformer hybrid network for detecting the breakage and separation degree of garlic cloves

Current research on garlic clove breaking primarily focuses on equipment development and optimization, with limited attention to elucidating the relationship between the breaking process and resultant damage or clove separation. To address this gap, this paper presents GarlicNet, a deep learning model that utilizes pressure signals generated during clove breaking to accurately detect breakage severity and separation extent. The model employs an enhanced adaptive channel attention mechanism to capture critical multi-channel information and a dual-branch attention structure to amplify salient features within pressure signals. Experimental results demonstrate high performance: GarlicNet achieved 95.6% accuracy, 95.9% recall, 95.9% precision, and 95.8% F1-score for breakage detection; corresponding values for separation detection were 96.5%, 96.5%, 96.5%, and 96.4%. Ablation studies confirm the efficacy of the Dynamic Multi-Channel Convolution Fusion (DMCF) module and Dual-Branch Attention Fusion (DAF) structure. Compared to benchmark models, GarlicNet exhibits superior detection performance and robustness. These findings validate the feasibility of predicting breakage and separation via pressure signal during clove breaking and underscore the model’s practical utility. This approach shows significant potential for mechanized garlic processing by reducing losses, improving efficiency, and advancing industrial automation.