The presence of infertile eggs and eggs with dead embryos in the incubator is one of the main problems in the hatchery. The existence of these eggs, in addition to occupying the space of the incubator, reduces hatching efficiency and causes contamination of the incubator and other eggs. Therefore, developing a non-destructive system to detect fertilized eggs, preferably before the start of the incubation process or day 0, can be an effective step in increasing the efficiency of the chicken industry and preventing the wastage of millions of infertile eggs, which can be consumed by humans if detected early. Until now, various methods have been developed to detect unfertilized eggs and dead embryos. In this thesis, the feasibility of using the hyperspectral imaging method for early detection of fertile eggs before (day 0) and after the start of incubation was investigated. For this purpose, an imaging chamber with effective light sources was designed and developed. Using a hyperspectral camera in the spectral range of 400–1000 nm, hyperspectral images of eggs were acquired before and during the incubation process. To prepare fertilized and unfertilized eggs with similar conditions in terms of hen’s age, standard diet, and maintenance management, 60 Leghorn laying breeder hens (Hy-Line W-36) and 4 roosters were purchased and kept in the farm of Isfahan University of Technology. The hens were randomly distributed into two sub-flocks: 30 hens without roosters and 30 hens with 4 roosters to prepare unfertilized and fertilizing eggs, respectively. Daily, samples were collected, and their hyperspectral images were acquired after measuring the physical characteristics. Subsequently, the samples were pro‎mp‎tly entered into the incubation device, and their hyperspectral images were captured on the first, second, third, and fourth days. Two methods of hyperspectral image data analysis, including spectral and pixel-based spatial analysis, were used for egg fertility measurement before and after the incubation process, as well as for the diagnosis of embryo development. In the spectral analysis, the background was first removed, and the spectral data belonging to the entire egg area (ROI) were extracted. The average spectrum of ROI was then calculated and preprocessed using various methods. This preprocessed data served as input for different classification methods, including soft independent modeling of class analogy (SIMCA), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), and artificial neural networks (ANN). In spatial analysis, the spectral information of 400 random points in hyperspectral images of each sample was used as input for ANN and deep learning methods. After confirming the performance of the classification models, these models were applied to the hyperspectral images of the test sample sets, and color images related to fertility and embryo development were extracted. The results of spectral analysis on day 0 showed that using the SIMCA method, the best classification accuracy (86.67%) was obtained with first and second derivative pre-processing. By examining the discrimination power plot, the wavelengths that had the greatest effect on the separation of the two classes were extracted. The spectral differences between the two pairs of selected wavelengths were used to simulate the first derivative pretreatment and were used as input to the ANN model. This model was able to separate fertilized and unfertilized eggs on day 0 with 93.33% accuracy. In diagnosing the process of embryo development using spectral analysis, the ANN model with first derivative preprocessing was able to detect embryo development on the first, second, third, and fourth days with accuracies of 96.10%, 96.10%, 94.81%, and 97.40%, respectively. In pixel-based spatial analysis, the best result in fertility detection on day 0 was obtained with first derivative preprocessing and the ANN method (95.83% accuracy in the test step).

احمد ميره اي , امين اله معصومي

محمد صدقي

مرتضي صادقي , سعيد انصاري مهياري , كاوه ملازاده