Abstract:
Abstract: Layer strains can be a particular preference for female chickens over males with respect to economic value. One-day-old male chicks are normally slaughtered once hatched in the incubation factories, leading to a huge waste of poultry resources, as well as raising serious concerns for animal welfare. It is since day 7 of incubation that the chicken embryo starts to feel pain. Thus, accurate identification can be used for the removal of the male eggs at the early stage of incubation, in order to effectively save cost and conform to the ethics for animal welfare. Generally, incubation is a complicated process concerning the inner biochemical activity and outer morphology evaluation, for example, the shape of the chicken embryo and the vessels. However, only a single information source was used, such as spectra or machine vision images, thus performing weakly in the recognition accuracy. This study aims to more accurately detect the gender of the chicken embryo at the early stage of incubation. A non-destructive decision fusion was proposed using both images and spectra using Random Forests (RF) and Dempster-Shafer (D-S) theory of evidence. A detection system was constructed to sample the transmission spectra and machine vision images of 566 chicken eggs. Day 4 of incubation was selected as the best detection time, while, the laying style of eggs was determined as placed horizontally. Machine vision images of the chicken embryo were treated via image processing, including morphological operations, and the Otsu algorithm. Then, the eleven texture features of the chicken embryo were extracted and scaled using grey level co-occurrence matrix, grey histogram statistics, and fractal dimension. The preprocessing method of spectra was determined as the scaling after the experiment. Four spectral features were extracted from the preprocessed spectra via Competitive Adaptive Reweighted Sampling (CARS). After that, two detection models were established using visual features and RF spectral. Five-fold cross-validation was then applied in the task of grid searching to optimize the two models. The machine vision model reached 78.00% accuracy with optimized parameters, while the spectral model was 82.67% accuracy for the test set. Furthermore, the feature fusion model was also constructed using texture and spectral features. The recognition accuracy of the test set only achieved 62.33% accuracy, indicating the mixed redundant information in the features. Finally, the decision fusion model was built via the D-S theory of evidence. The basic probability assignment functions were obtained from the optimal RF models of images and spectra. Then, the decision fusion model was established using the fusion principle and threshold of the D-S theory of evidence. Consequently, the fusion model reached 88.00% accuracy, particularly with 90.00% and 86.25% accuracy for the female and male eggs. Besides, 2.843 s was used for the D-S model to detect each egg. Anyway, the decision fusion can be expected to realize the gender detection of the chicken embryo at the early stage with a higher accuracy than before. The finding can provide a potential solution to the commercial application in the poultry industry.