Abstract:The presence of cracks reduces egg quality and safety, and can easily cause food safety hazards to consumers. Machine vision-based methods for cracked egg detection have achieved significant success on in-domain egg data. However, the performance of deep learning models usually decreases under practical industrial scenarios, such as the different egg varieties, origins, and environmental changes. Existing researches that rely on improving network structures or increasing training data volumes cannot effectively solve the problem of model performance decline on unknown egg testing data in practical egg production. To address these challenges, a novel and robust detection method is proposed to extract max domain-invariant features to enhance the model performance on unknown test egg data. Firstly, multi-domain egg data are built on different egg origins and acquisition devices. Then, a multi-domain trained strategy is established by using Maximum Mean Discrepancy with Normalized Squared Feature Estimation (NSFE-MMD) to obtain the optimal matching egg training domain. With the NSFE-MMD method, the original deep learning model can be applied without network structure improvements, which reduces the extremely complex tuning process and hyperparameter adjustments. Finally, robust cracked egg detection experiments are carried out on several unknown testing egg domains. The YOLOV5 (You Only Look Once v5) model trained by the proposed multi-domain training method with NSFE-MMD has a detection mAP of 86.6% on the unknown test Domain 4, and the YOLOV8 (You Only Look Once v8) model has a detection mAP of 88.8% on Domain 4, which is an increase of 8% and 4.4% compared to the best performance of models trained on a single domain, and an increase of 4.7% and 3.7% compared to models trained on all domains. In addition, the YOLOV5 model trained by the proposed multi-domain training method has a detection mAP of 87.9% on egg data of the unknown testing Domain 5. The experimental results demonstrate the robustness and effectiveness of the proposed multi-domain training method, which can be more suitable for the large quantity and variety of egg detection production.

Egg Weight Detection on Machine Vision System

A Computer Vision-Based Automatic System for Egg Grading and Defect Detection

Study on the Model of Egg Weight Detecting Based on its Geometry

Image Classification of Chicken Embryo Based on Matched Filter and Skeleton Curvature Feature

Egg Freshness on-Line Estimation Using Machine Vision and Dynamic Weighing

Prediction of Shelled Shrimp Weight by Machine Vision

Estimating yolk weight of duck eggs using VIS-NIR Spectroscopy and RGB images and whole egg weights

A Novel Method for Measuring the Volume and Surface Area of Egg

Monitoring storage time and quality attribute of egg based on electronic nose.

Visual detection of eggs based on deep learning for egg picking robot

Improvement of Accuracy for Egg Dynamic Weighing Using Artificial Neural Network

Robust Detection of Cracked Eggs Using a Multi-Domain Training Method for Practical Egg Production

Separate weighing of male and female broiler breeders by electronic platform weigher using camera technologies

Prediction of Egg Storage Time and Yolk Index Based on Electronic Nose Combined with Chemometric Methods

An automatic sorting system for unwashed eggs using deep learning

Detection of Egg Freshness Using Electronic Nose

Study on egg sorting model based on visible-near infrared spectroscopy

Application of Optimized Digital Filters and Asymmetrically Trimmed Mean to Improve the Accuracy of Dynamic Egg Weighing

Electronic Nose Technique Potential Monitoring for Storage Time and Quality Attribute of Egg

Development of a machine vision-based weight prediction system of butterhead lettuce (Lactuca sativa L.) using deep learning models for industrial plant factory

Nondestructive Detection of Egg Freshness Based on Infrared Thermal Imaging