Lightweight bucket tooth detection network based on AKConv

As a key vulnerable component of mining mechanical front shovel excavators, the wear condition of bucket teeth directly affects equipment operational efficiency and safety. However, due to complex working environments where bucket teeth are prone to occlusion, adhesion, mud coverage, and variable lighting conditions, manual inspection or traditional image processing-based detection methods exhibited low efficiency and high error rates, failing to meet the demands of intelligent construction management. Therefore, researching on deep learning-based object detection for bucket teeth holds significant theoretical and practical value. This study proposes a bucket teeth detection method suitable for limited data samples, integrating adjustable kernel convolution（AKConv）, the YOLOv12 network, and a data augmentation algorithm. First, the labeled original bucket teeth dataset is augmented with a 1∶1 random data enhancement strategy to generate diversified images and labels, simulating bucket teeth detection scenarios under varying imaging conditions and improving dataset diversity and robustness. The augmented dataset is then fed into the improved AKConv-YOLOv12 network for 100 training iterations, achieving a precision of 97.2%, recall of 96.7%, and mAP（mean Average Precision） of 98.0%, with a model size of 4.8 MB. Compared to the original YOLOv12 network, the precision and recall decreases by 1.0 and 0.9 percentage points, respectively, while the mAP drops by 0.6 percentage points. However, the model size is reduced by 11.1%. Finally, tests on both augmented and non-augmented datasets under varying occlusion levels demonstrated detection confidence levels consistently around 0.8. The results indicate that the proposed small-sample detection method, combining data augmentation and AKConv-YOLOv12, effectively addressed the limitations of homogeneous data samples, ensuring high robustness and stable detection performance. Moreover, while maintaining nearly equivalent accuracy, the model size is significantly reduced, enhancing portability. This approach provided essential technical support for research on mining machinery and excavator maintenance.