Unsupervised delineation of chicken eggs based on the universal avian egg model

Abstract Due to advancements in computing technology, egg image processing has become an industrial standard in poultry research. It has been utilized at various stages of chicken breeding and hatchery, as well as egg quality assessment and control. Thus far, existing methods are impeded by the need for problem specific provisions, whose inherent setting is counterintuitive, or those involving overly complex artificial neural networks. Motivated by the latest development of universal avian egg models, this paper proposes an automated egg delineation method, based on expectation-maximization framework. Accordingly, only a few model and pose parameters were required, discarding the need for large, annotated training dataset or excessive priors. Moreover, numerical treatments were also implemented to ensure its stability in practice. The experiments on a public dataset demonstrate that the method could segment eggs of various shapes and sizes and be acquired under different perspectives and lighting conditions. Reported performance analyses revealed the average dice coefficient and intersection of union of 0.9782 and 0.9575, respectively. Its resilience to noise and partial occlusion were also evaluated. Thus, the proposed method not only enhances this critical task, but is also expected to be integrable with existing egg image recognition workflows in precision farming and food inspection. • Computerized implementation following the latest development on universal avian egg models. • It imposes minimal constraints on breeds, sizes and shapes, and image acquiring conditions of egg samples. • It does not require any model training on large pre-annotated dataset, minimizing its computational footprint. • Validations performed on a publicly available dataset allows reproducibility and fair future comparative studies. • The resulting delineation offers accurate collections of biological attributes present within an egg.