Purpose To solve the problem of misidentification of broiler behavior caused by complex and changeable background, aggregate occlusion, multi-scale targets and motion blur, proposing a recognition model of broiler behavior throughout the full growth cycle based on the improved YOLO11s (YOLO11s-Broiler).
Methods By introducing the sobel-edge feature aggregation module (SEFA) and CSP-dualconv feature refinement module (DCFR) to optimize the backbone network, using the ADown module to replace the sampling module under the strided convolution, a double calibration feature pyramid network (DCFPN) structure including repconv multi-stage feature aggregation (RMSFA), context feature calibration (CFC) and spatial feature calibration (SFC) modules was designed to build a broiler behavior recognition model. Seven behavioral image data of caged broilers, such as exploration, panting, and grooming, were collected throughout the full growth cycle, and the data set was expanded. The improved model was used to conduct experimental analysis on broiler behavior.
Result After the introduction of modules such as SEFA and CSP-DCFR, the mean average precision (mAP) had been improved, with a maximum increase of 2.2%. The mAP values of the improved model in the brooding period, rearing period, fattening period and comprehensive test set reached 93.9%, 89.2%, 89.2% and 90.9%, respectively, which were 2.6%, 2.6%, 2.0% and 2.2% higher than the baseline model YOLO11s, respectively. Compared with the other eight target detection models, the improved model had the highest mAP value, and the computational complexity and number of parameters were obviously reduced.
Conclusion The improved YOLO11s model has the advantages of high accuracy, strong generalization and good robustness, and it can provide a useful reference for behavioral recognition of poultry throughout their growth cycle in complex environments and for modern breeding management.
DownLoad: