Detection of mare parturition through balanced multi-scale feature fusion based on improved Libra RCNN

Once a mare experiences parturition abnormalities, the outcome between a live foal and a stillborn can change rapidly. Automated detection of mare parturition and timely human intervention is crucial to reducing risks during mare and foal parturition. This paper addresses the challenges of manual monitoring of parturition in large-scale equine facilities due to the unpredictability of mare parturition timing, proposing an algorithm for detecting mare parturition through a balanced multi-scale feature fusion based on an improved Libra RCNN. Initially, a ResNet101 backbone network incorporating the CBAM attention module was used to enhance parturition feature extraction capability; subsequently, a balanced content-aware feature reassembly feature pyramid, CARAFE-BFP, was employed to mitigate data imbalance effects while enhancing the quality of feature map upsampling; finally, the GRoIE module was utilized to merge CARAFE-BFP’s multi-scale features, improving the model’s perception of multi-scale objectives and minor feature changes. The model achieved a mean average precision of 86.26% in scenarios of imbalanced positive and negative samples of mare parturition data, subtle parturition feature differences, and multi-scale data distribution, with a detection speed of 15.06 images per second and an average recall rate of 98.17%. Moreover, this study employed a statistical method combined with a sliding window mechanism to assess the algorithm’s performance in detecting mare parturition in video stream continuous monitoring scenarios, achieving an accuracy rate of 92.75% for mare parturition detection. The algorithm proposed in this study achieved non-contact, stress-free, intensive, and automated detection of mare parturition, also demonstrating the immense potential of artificial intelligence technology in the field of animal production management.