Cigarette detection in diverse environments using YOLO-CD with adaptive multiscale feature extraction

Cigarette detection plays a crucial role in environmental protection and public health. However, current deep learning based algorithms for small object detection often suffer from poor accuracy and are highly susceptible to environmental interference and variations in lighting conditions. To address these challenges, this paper introduces an end-to-end cigarette detection algorithm: YOLO-CD. First, a novel adaptive multi-scale feature extraction module is designed to more comprehensively capture features across different scales. Second, to mitigate the distortion and loss of small object information in deep networks, a more flexible and efficient downsampling design is employed. Additionally, a deep supervision mechanism is introduced by adding auxiliary detection branches in the intermediate layers of the network, effectively enhancing the model's ability to capture small object features without increasing computational complexity. Experimental results demonstrate that YOLO-CD outperforms existing mainstream methods on public datasets, achieving a 3.0% improvement over the baseline model while also reducing the number of parameters and computational load.