Objective Due to the lush vegetation and disorderly trees in the wild environment, as well as the influence of factors such as environment, weather, and lighting, infrared cameras are prone to triggering shooting errors, resulting in the capture of a large amount of waste film, which requires a lot of manpower for screening. To solve such problems, based on the YOLOv7 model, this paper has made lightweight improvements to achieve automatic screening of waste pieces.

Method This study constructed a dataset of 2 172 wildlife images collected from the Beijing Wuling Mountain Nature Reserve in the period of 2014−2015, and marked the positions of animals in the images. YOLOv7 network was improved in different ways. MicroBlock was introduced to replace the backbone network of YOLOv7, and the SPPCSPC structure was light-weighted to reduce the model parameters. SIoU loss, LNDown downsampling, and BiFPN were used to improve the model’s ability to detect animals. YOLOv5-m, YOLOv5-l, Ghost-YOLOv5-l, YOLOv6, YOLOX-M, and YOLOR-CSP models were trained on an Snapshot Serengeti camera trap subset dataset containing 10 000 images, and the screening effects of the model on wildlife images were compared. Transfer learning was used to train a self-built wildlife dataset, and the training effects of freezing different layers was tested.

Result The improved model based on YOLOv7 reduced inference time by 14.3%, floating-point operations per second by 33.5%, and parameters by 17.8% compared with the YOLOv7 network. The error detection of the improved YOLOv7 model was also better than that of YOLOv7. Although the improved YOLOv7 did not achieve the best performance in all indicators compared with other models, it achieved a better balance between detection time and accuracy. In the self-built dataset, the unfrozen weight method had the best effect, and average precision was 12.6% higher than that of the model without transfer learning.

Conclusion This study provides a faster and more accurate screening solution for wildlife monitoring networks in the Mountain area of Beijing Miyun.