Pig Target Detection from Image Based on Improved YOLO V3

Smart farming has always been one of the current research hotspots. Reflected by the behaviors and moves, the physiological conditions of pigs can be detected. The inability to detect the behaviors of pigs at scale has become the urgent issues. This makes the recognition of pigs an extremely significant problem. Building on the prior work on picture-based recognition of target detection, this paper put forward an improved YOLO V3 to detect pigs from image. To overcome the lack of pig’s pictures training data, transfer learning is used. To improve the accuracy of algorithm, attention mechanism is introduced into YOLO V3. Results show the algorithm we exploited can efficiently complete the task for pig real-time detection. Compared with the classical YOLO V3, the improved YOLO V3 has better metrics on precision, recall, F1 score and average precision. The improved model achieves result: 94.12% AP. The result is encouraging enough to make people collect more labeled pig’s picture data to improve the generalization capability of the algorithm.