Skip to main content
ARS Home » Southeast Area » Stoneville, Mississippi » Crop Production Systems Research » Research » Publications at this Location » Publication #366646
Title: Evaluating convolutional neural networks for cage-free floor egg detection

Author
item LI, GUOMING - Mississippi State University
item XU, YAN - Mississippi State University
item ZHAO, YANG - Mississippi State University
item DU, QIAN - Mississippi State University
item Huang, Yanbo

Submitted to: Sensors
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/5/2020
Publication Date: 1/7/2020
Citation: Li, G., Xu, Y., Zhao, Y., Du, Q., Huang, Y. 2020. Evaluating convolutional neural networks for cage-free floor egg detection. Sensors. 20:332. https://doi.org/10.3390/s20020332.
DOI: https://doi.org/10.3390/s20020332

Interpretive Summary: Automatic detection of floor eggs in cage-free hen housing is expected to reduce labor and improve the management. Scientists from Mississippi State University and USDA ARS Crop Production Systems Research Unit at Stoneville, Mississippi have collaboratively evaluated three convolutional neural networks through computer vision for effective detection of the floor eggs. Among the three network models the favorite one was selected and assessed under different hen housing conditions. The results indicated that the favorite model can accurately detect floor eggs in cage-free housing environment and the method developed in this research can be used to develop a tool for automatic floor egg detection and collection in practice.

Technical Abstract: Automatic detection for floor egg in cage-free (CF) hen housing relies on robust computer modeling tools and vision systems. The objectives of this study were to 1) select a favorable model built by Convolutional Neural Network (CNN) to detect floor eggs through a comparative evaluation of three typical object detection CNNs, single shot detector (SSD), faster region-based CNN (faster R-CNN), and region-based fully convolutional network (R-FCN) under a typical CF housing environment; 2) investigate the performance of the favorable CNN model for floor egg detection under different settings reflecting a range of common CF environments and conditions; and 3) test the generalizability of the favorable CNN model under random settings. The results indicated that SSD model had the fastest processing speed (74.6 ms·image-1) but the lowest recall (82.4%) and accuracy (82.3%) among the three models. The R-FCN model had the slowest processing speed (239.7 ms·image-1) and the lowest precision (90.2%). The faster R-CNN model had the highest precision (99.9%), recall (100.0%), and accuracy (99.9%) and a medium image processing speed (180.6 ms·image-1), thus was selected as the favorable model for floor egg detection. The faster R-CNN detected the floor eggs almost perfectly under different CF environments and conditions, except for brown egg under 1-lux light intensity. When tested under random settings, the faster R-CNN model had 88.8-93.1% precision, 99.0-100.0% recall, and 88.8-92.2% accuracy for detecting white or brown eggs. It is concluded that a properly selected CNN model may accurately detect floor eggs in CF housing environment and has a great potential to serve as a useful detection tool for automatic floor egg collection systems.