An underground coal mine multi-target detection algorithm

Currently, underground coal mine target detection algorithms based on deep learning show poor performance in detecting complex small targets under conditions of uneven light intensity distribution, complex target environments, and imbalanced multi-class target scale distribution, often resulting in missed detection and false detection. To address these issues, based on the single-stage target detection algorithm YOLOv8n, this study proposed an underground coal mine multi-target detection algorithm based on feature extraction by dynamic snake convolution (FEDSC)-feature fusion by bi-directional feature pyramid network and semantic and detail fusion (FFBD). FEDSC replaced the backbone network of YOLOv8n to expand the receptive field, while FFBD acted as the neck network to reduce target false detection and missed detection. Additionally, a decoupling detection head of SIoU was used as the detection layer to improve the model's adaptability to small targets and the convergence speed. The results showed that: ① The [email protected] of the FEDSC-FFBD algorithm was 97.00%, the number of model parameters was 4.22×106, and the number of floating point operations per second was 21.7×109. ② The [email protected] of the FEDSC-FFBD alorithm was 3.40% higher than the YOLOv8n algorithm, and the recognition accuracy of the helmet small target was 90.90%, 11% higher than the YOLOv8n algorithm. ③ Compared with other YOLO series algorithms, the FEDSC-FFBD algorithm achieved the highest [email protected], which was 3.60%, 1%, 10.50%, and 6.40% higher than YOLOv5s, YOLOv9c, YOLOv10n, and YOLOv11n algorithms, respectively. ④ The FEDSC-FFBD algorithm improved the detection accuracy of multi-class targets and reduced missed detection and false detection of small targets under conditions of uneven light intensity distribution, complex target environments, and imbalanced target scale distribution in underground coal mine. The underground coal mine multi-target detection algorithm based on FEDSC-FFBD overcame the challenge of small-scale target detection caused by uneven light intensity distribution without relying on image quality enhancement algorithms.