<![CDATA[Accurate early detection of egg fertilisation status is necessary for effective incubation management in chicken production in order to avoid energy waste and decreased hatchery productivity brought on by infertile or non-viable eggs. Due to their comparable perceptual traits, conventional candling inspection relied on manual observation, which introduced subjectivity and made it challenging to distinguish between fertilised and blighted eggs early on. This study suggested an automated multistage fertilisation prediction method based on candling image analysis, utilising a convolutional neural network framework to get around this restriction. Rather than using traditional binary classification, the suggested system allowed for progressive monitoring of embryonic growth. On incubation days 1, 7, 14, and 21, candling photos were taken from native chicken eggs and classified into three groups: fertilised, infertile, and blighted. To enhance feature extraction efficiency under constrained dataset conditions, a transfer learning technique utilising the MobileNetV2 architecture was implemented. To guarantee consistent learning performance, image preprocessing, augmentation, model training, and validation were carried out. Precision, recall, F1-score, and classification accuracy were used as assessment measures. According to experimental findings, the suggested model produced consistent classification results for both fertilised and infertile eggs, with validation accuracy ranging from 90 to 95% throughout the incubation period. The results of multistage prediction showed consistent decision-making throughout the observation of embryo development. However, during intermediate incubation stages, visual uncertainty with fertilised eggs led to decreased performance in recognising blighted eggs. All things considered, the suggested method showed great promise as a nondestructive intelligent system for early fertilisation prediction. To increase the accuracy of blighted egg classification, more dataset expansion and model improvement were still required.]]>
