NONDESTRUCTIVE TESTING OF DUCK EGGS DURING INCUBATION USING YOLO-LITE

At present, elimination of eggs with dead embryos in China's poultry-egg incubation industry mainly relies on manual inspection and a lack of relevant automatic-detection equipment restricts seriously the development of egg-incubation industry. Accurate identification of dead eggs is a key technical task for solving this problem. In this study we design a set of image-acquisition devices for analyzing duck eggs on the stage when they are removed from the incubation tray (25 days after incubation beginning). We suggest improved object-detection algorithm based on a YOLOv4 network and identify dead embryos with high accuracy during the egg-hatching period. According to the characteristics of collected images of the breeding-duck eggs, we remove the head subnetwork for detecting small objects in the YOLOv4 network and simplify the backbone subnetwork of this network to improve the detection efficiency. The experimental results testify that the average recognition accuracy of our YOLO-Lite network is equal to 98.33%, the recall rate amounts 94.12% and the single-frame image-recognition time is about 15 ms. These figures are better than the corresponding parameters 96.67%, 88.89% and 32 ms, which are typical for the technique before our improvement. Therefore our results can provide a basis for the further research and development appropriate intelligent detection equipment.