A multi-target cow face detection model in complex scenes

The development of intelligent agriculture has accelerated the automation and scale of cattle farming. Recognizing cattle faces as a significant biological characteristic is crucial for accurate reproduction and health tracking. We propose a multi-target cow face detection model (MT-CF-DM) in complex scenes based on the YOLOv7 framework. The backbone of our proposal model consists of the GhostNet and the CBAM (convolutional block attention module) attention mechanism to improve the model’s perception of various scales and features. Additionally, we propose three novel modules of the neck network: SPPFCSPC (Spatial Pyramid Pooling-Fast Cross-Stage Partial Connection), a novel BiFPN, and C2f (CSP Bottleneck with 2 convolutions). SPPFCSPC is to reduce the number of model parameters. New BiFPN is to utilize feature information and improve the model’s detection capability. C2f is to maintain the network’s lightweight nature and obtain more comprehensive gradient flow data. Moreover, a method for establishing cow face datasets is proposed based on the MT-CF-DM. During the testing phase, our model combines with a cropping module to simultaneously acquire a dataset for cow faces. The experimental results show the parameter of the proposal model is 15.8 M, the model size is 22.3 M, GFLOPs are 25.8, FPS is 97.6, mAP is 98.5%, precision is 98.8%, recall is 97.3%. Compared to YOLOv7, our proposed model reduces parameters by 56.8% and has a 48.3 increase in FPS. The model demonstrates high accuracy and small number of parameters for detecting cow faces in complex environments, suitable for edge computing system applications.