Shell thickness of brown eggs: Proposal and implementation of an automated method for analysis and classification

Abstract

• An optimized computer vision system for recognizing acceptable and unacceptable shell thickness in brown eggs. • Automation of the new shell thickness vision method for rapid classification of excitants. • Proposal of a prototype for eggshell thickness classification that is less costly to purchase and maintain. Although computer vision and artificial intelligence solutions exist for automated egg classification, most poultry farmers in Cameroon still use manual sorting based on visual assessment of eggshell characteristics. Existing automated systems remain largely unsuitable for this context due to their high cost, bulky design, significant energy consumption, and the need for skilled personnel for operation and maintenance. To address these limitations, this study proposes a computer vision and artificial intelligence architecture that exploits optical properties to analyze the red, green, and blue components captured by a camera when observing brown eggshells. Five subclasses of shell thickness are defined: very good thickness, good thickness, medium thickness, poor thickness, and very poor thickness. Six classification features are extracted from the histograms of eggshell images using two intensity intervals. A dedicated imaging device consisting of an egg holder, a light-emitting diode illumination system and, a camera was designed to capture the images. The dataset is collected from the poultry farm of the Faculty of Agronomy and Agricultural Sciences (FASA)\Cameroon. Several machine learning algorithms are evaluated, and the Random Forest model achieves the best performance with an accuracy of 98.3%, an F1-score of 97%, a recall of 97%, and a precision of 96%. The architecture is implemented on a microcomputer to meet the operational constraints of Cameroonian poultry farms. Tests conducted at the FASA farm demonstrate a startup time of four minutes, an average prediction time of 17.5 seconds, and a prediction rate of 100% for the good/poor thickness classification on the test sets.