Non-destructive Identification For Gender Of Chicken Eggs At The Early Stage Of Incubation Based On Spectroscopy And Machine Vision

For poultry breeding,males of layer strains can product neither eggs nor as much high-quality meat as that of broiler strains.There is so little economic value in males that one-day-old male chicks are slaughtered or poisoned once hatched,which means that half of the total hatchlings are wasted,raising more and more concern for animal welfare.Therefore,identifying the gender of chick embryos at the early stage of incubation and removing excess male eggs can not only cut down the waste of the hatching process,but also improve economic benefits.Chicken eggs were used as experiment materials in this research.Two detection systems were established based on visible-near infrared(VIS-NIR)spectroscopy and machine vision respectively.The feature parameters indicating gender information were extracted from the raw spectra and raw images.After,the machine learning models based on single information source and fusion information were established respectively.At last,a detection software using multi-source information was developed to realize the non-destructive gender detection of chicken embryo during incubation.The main contents and conclusions of this research are as follows:(1)The construction and optimization of the detection systems.Hardware selection was conducted for two sets of detection systems.Then the detection systems were established,the parameters of which were adjusted properly through experiments.(2)The establishment of the gender detection model using spectroscopy.The VIS-NIR spectra of chicken eggs were sampled during incubation.Through comparing,the day 4 of incubation was selected for modeling and eggs were placed horizontally when sampled.When using a combination of two wavelengths of the transmission spectrum,there was a significant difference(P<0.01)between males and females in T575.02/T610.00 and other ratios at the day 4 and day 7 of incubation.After comparing the models,the preprocessing method was determined as normalization.Competitive Adaptive Reweighted Sampling(CARS)and Successive Projection Algorithm(SPA)were used to extract the features for modeling.Also,the T575.02/T610.00 ratio or other ratio was merged with the selected features respectively to build feature fusion models.The results suggested that the model using the four features selected by CARS had the best performance and could replace the model built with the raw spectra.Moreover,the spectral detection models were established using Random Forest(RF),Gradient Boosted Decision Trees(GBDT)and Support Vector Machine(SVM)respectively.In terms of the model’s accuracy and Area Under Curve(AUC),the RF model was determined as the best and then optimized via grid searching.According to the experiment,when the number and the max depth of trees were set as 16 and 90,the RF model reached 82.67% accuracy,with 91.42% accuracy for females and 75.00% for males respectively.(3)The establishment of the gender detection model using machine vision.The machine vision images of chicken eggs were sampled during incubation.To acquire high-quality images of chicken embryo for feature extraction,the day 4 of incubation was selected for modeling and the eggs were placed horizontally when sampled after repeated experiments.A series of image processing methods were utilized to obtain chicken embryo images from the raw images,including adaptive cropping,image denoising,image enhancement,threshold segmentation,image mosaic and mapping.Then,eleven texture features were extracted using Gray-Level Histogram Statistics(GLHS),Gray-Level Co-occurrence Matrix(GLCM)and fractal dimension(FD).Moreover,the two features with the highest importance were selected from the eleven features for modeling.The parameters were optimized via grid searching.When the number and the max depth of trees were set as 70 and 48,the model reached its best performance at 78.00% accuracy,with 81.43% accuracy for females and 75.00% for males.(4)The establishment of the information fusion model using spectroscopy and machine vision.In order to integrate the two types of incubation information of chicken eggs,feature fusion and decision fusion methods were used respectively to build detection models for gender detection.The results suggested that the decision fusion model based on D-S theory of evidence reached 88.00% accuracy in gender detection,with 90.00% accuracy for females and 86.25% for males.Combining different information sources,the D-S model could make more comprehensive and accurate decision,showing good performance in detecting genders of chicken eggs at the day 4 of incubation.(5)The development of chicken egg gender detection software.A type of gender detection software was developed based on the proposed models in this paper.The software allows users to input the files of spectroscopy and machine vision and select the detection model.Also,the prediction results can be visualized in this software.The results indicated the feasibility of the D-S fusion model based on spectroscopy and machine vision,providing a new method for gender determination of chicken eggs during incubation.