A novel daily behavior recognition model for cage-reared ducks by improving SPPF and C3 of YOLOv5s

Abstract

Intensive duck farming can improve production efficiency and reduce environmental pollution. In modern intensive farming, ensuring the well-being and health of ducks is a paramount concern. Generally, the health status of ducks is determined by monitoring their daily behaviors, such as eating. However, research on cage-reared duck daily behavior recognition is scarce. Therefore, this study proposes a cage-reared duck daily behavior recognition model based on the improved YOLOv5s, denoted DBR-YOLOv5s for notational convenience. Specifically, to tackle the interfered features caused by the duck cage, an improved shrinkage mechanism is introduced in block SPPF of YOLOv5s. To decrease the feature information loss, the convolution operation replaces the maxpool operation in SPPF. Moreover, to cope with the issue of occlusion, the last three C3 blocks in module Neck of YOLOv5s are optimized via the multi-scale convolution operation, promoting the capability of DBR-YOLOv5s to extract contextual information. Extensive experiments were conducted on the self-constructed duck daily behavior dataset. The test precision of the proposed DBR-YOLOv5s is 92.8% for drinking, 96.8% for lying, 93.8% for standing, 98.5% for eating, 89.9% for preening, and 96.7% for spreading. Compared with the state-of-the-art YOLOv8x and YOLOv9c models, the average precision of DBR-YOLOv5s is 1.3% and 2.4% higher, respectively. The results indicate that the proposed DBR-YOLOv5s is effective for cage-reared duck daily behavior recognition, providing a non-contact method for duck daily behavior recognition.