An improved YOLOv5 model: Application to leaky eggs detection

Abstract

In order to quickly and correctly detect the leaky eggs on the egg processing line, this article proposed a leaky eggs detection model based on an improved YOLOv5 network. Firstly, we replaced the feature pyramid network and path aggregation network (FPN + PANet) with bidirectional feature pyramid network (BiFPN) to improve the multi-scale object recognition ability without increasing computational costs and improve the detection accuracy of small objects in egg images. Secondly, we added the Convolutional Attention Mechanism Module (CBAM) to enhance the ability for perceiving target information to improve the model's detection performance. This result shows that the improved model for detecting leaky eggs has an average precision (AP) of 92.4%, which is 3.9% higher than the original network and also outperforms other target detection models. Moreover, in three group of comparison experiments completed on a laboratory platform with different conveying speeds, the detection accuracy of leaky eggs was 97.1%, 91.4% and 87.1%, respectively, indicating that the improved YOLOv5 model has good performance but the missed rate will increase with the increase of conveying speed. Finally, the test results on the factory egg processing line (Kyowa/FP-600) show that the improved model can detect the leaky eggs appearing in the visual window in real time, meeting the detection requirements of the egg processing line.