The laying duck behavior pattern is an important indicator for assessing the health and welfare status of ducks in cage farming. An object detection algorithm based on improved YOLO v4 (you only look once) was proposed to identify multiple behavior patterns in laying ducks by machine vision, and the different behavior patterns provided a basis for duck breeding management scheme. By replacing the backbone feature extraction network MobileNetV2 and using the depthwise separable convolution, this algorithm can improve the detection accuracy while reducing the number of model parameters and effectively improving the detection speed. The parameter-free attention mechanism SimAM module was introduced in the prediction output part to further improve the model detection accuracy. By using this algorithm to detect the cage-reared laying duck behavior validation set, the mAP value of the optimized model reached 96.97% and the image processing frame rate was 49.28f/s, which improved the mAP and processing speed by 5.03% and 88.24%, respectively, compared with the original network model. Comparing the effect with commonly used object detection networks, the improved YOLO v4 network improved the mAP values by 12.07%, 30.6% and 2.43% compared with Faster R-CNN, YOLO v5 and YOLOX, respectively. The improved YOLO v4 network proposed was experimentally studied. The results showed that this algorithm can accurately record the behaviors of cage-reared ducks at different time periods, helping identify abnormal conditions of ducks according to the different behavior patterns exhibited by ducks, such as some behaviors occurring for abnormal periods of time or during abnormal periods. The research result can provide valuable guidance for duck breeding management and enable technical support for implementing automated and intelligent management of duck houses.