Development of a new grading system for quail eggs using a deep learning-based machine vision system

Quail egg size varies due to differences in age, nutrition, and genotype of birds. The grading of eggs is significant, particularly for quail breeding farms, as it facilitates pairing similar phenotypic features for selecting optimal genetic strains. However, this process currently requires skilled manpower and complex equipment. In this work, we introduce a new system to grade quail eggs according to their size using a deep learning-based machine vision approach. The prototype was able to classify and segregate eggs into up to 4 classes: small, medium, large, and extra large. The new system was divided into two modules: the classification module and the sorting module. In the first module, each egg was tracked by the deep SORT algorithm and graded proportionally to the bounding box sizes using the deep learning-based computer vision models; in the second module, a pneumatic system was developed along with a damage prevention designed conveyor belt for sorting the egg classes. The classification accuracy rate found from comparing real classes and machine vision predictions showed that 4 or 3 classes were reliable in sorting out egg sizes with higher confidence, proving to be useful at a feasible implementation cost to support farmers. Three object detection algorithms were compared: EfficientDet, CenterNet, and YOLOv7. The models were scaled into 2 network sizes, 512 x 512 and 1024 x 1024, and the results were compared in terms of egg class prediction accuracy and real-time inference speed. In terms of class prediction accuracy, the best results for grading quail eggs into 4 classes were achieved by the EfficientDet-1024, CenterNet-512, and YOLOv7-1024 models, which correctly classified 78 %, 77.5 % and 72 %, respectively. While classifying the eggs into 3 classes, again, the best model was observed from EfficientDet-1024 (86 %), followed by CenterNet-512 (84 %) and YOLOv7-1024 (80 %). Regarding the trade-off between real-time inference speed and grading accuracy, YOLOv7 outperformed the compared models by inferencing at 40.38 % faster than CenterNet-512, which was the second fastest and most accurate grader. As result, the best compared model was estimated to grade at pace of 797 quail eggs per hour. The proposed combined machine vision-based system can be further scaled up either in small-scale farming or industrial anticipation for consumer satisfaction and to ensure safe production traceability.