Abstract: Abstract: The quality of breeding duck eggs dominates the hatching rate and healthy hatchling rate. The selection of high-quality breeding duck eggs can effectively reduce the occupied hatching resources and the wastes that are caused by the hatching failure. It is a high demand to timely screen out the breeding duck eggs unsuitable for hatching for other purposes. In this study, a novel detection system was designed to screen the external quality of breeding duck eggs before hatching using image processing and deep learning. The light source was arranged below the track. The shooting of the industrial camera was controlled by the photoelectric sensor. Three transmission images of the duck eggs were then collected, when the duck eggs entered the dark box. Furthermore, the real major axis was calculated by the maximum major axis of the duck eggs in these images, whereas, the real minor axis was also calculated by the average minor axis of the duck eggs. The calculated major axis value was then divided by the minor axis value to obtain the egg shape index. At the same time, the duck egg image was flipped in the long-axis direction, and then overlapped with the original image. The ratio of the non-overlapping part to the duck egg area was calculated to obtain the crisscross value, which was used to measure the similarity of the duck egg heads. The dirt feature was extracted to calculate the dirt area using the threshold segmentation in the HSV color space. The reason was that the color of the dirty part was generally darker than that of the crack on the duck egg surface in the transmission image. More importantly, the accuracy of threshold segmentation depended mainly on the different brightness of the duck egg in the transmission image, due to the difference in eggshell color and thickness. Therefore, the gamma transform was used to balance the color and brightness of the image before the HSV threshold segmentation. As such, the brightness reached a similar range in the image. Meanwhile, the adaptive threshold segmentation was used to preprocess the image sample data. The crack information was extracted in the case of uneven brightness in the image. The previously obtained image with the dirt feature was used as the mask to filter the crack feature image, in order to remove the interference caused by the dirt information. Finally, the convolution neural network (CNN) EfficientNetV2 was improved to establish the crack identification network CrackNet, in order to identify the cracks of dirty duck eggs. The inspection system was then obtained for the external quality of breeding duck eggs. Consequently, the size, egg shape index, and dirt area of breeding duck eggs were calculated to evaluate the similarity between the two ends of duck eggs, particularly for the detection of surface cracks. The mean square error (MSE) of the calculated long axis and egg shape index was only 0.220 2 mm2 and 0.000 058, respectively. The crack identification accuracy of CrackNet was 98.03%. The finding can also provide technical support to screen the breeding ducks before hatching.