Comparison of SSD Models for Fertile Egg Image Classification

Breeding laying hens and eggs products are prevalent in many parts of Thailand. Egg produced from these farms are important economic product. One important process in industrial incubation is the culling out of eggs that are unlikely to hatch before setting them into an incubator. In this paper we propose a novel method for real-time egg image classification using Single Shot Detector (SSD) with Transfer Learning from MobileNetV2 and RestNet101 model built from Microsoft COCO dataset. Images of eggs on its side are collected, cleaned, augmented, and labeled. The results of predictions using our two models are similar with mAP@.50IoU at 0.99, however SSD MobileNetV2 is faster to train. A preliminary test of models on images fed from a webcam show good performance with images with one egg, but still need improvement when there are 4 eggs in one image.