It is one of difficult problems to be resolved in egg hatching industry to identify the fertile information of hatching eggs and eliminate infertile eggs prior to the incubation. Many infertile eggs have been wasted in the process of incubation every year, which has caused considerable economic loss. The existing domestic infertile egg detection mainly depends on traditional manual candle method. However, this detection method requires high intensity of labor and is time-consuming. In addition, the result of detection is subjective and its accuracy can not be guaranteed. The detection of infertile eggs prior to incubation can improve the economic efficiency of incubation and the quality of egg processing in late period, and it can bring considerable economic benefits. This paper proposed that the hyperspectral imaging technology consisting of image and spectral information and the relevance vector machine (RVM) were used for detecting the fertile information of eggs before incubation. To build a hyperspectral transmission image acquisition system, the light source, the light intensity, the resolution, the exposure time, the platform moving speed and other parameters were adjusted when the images of hyperspectral instrument were captured. Ultimately, the exposure time of the camera was determined as 0.1 s, the resolution of image as 400×400 pixels, and the platform moving speed as 1.7 mm/s. Before hatching eggs incubation, hyperspectral images system was used to acquire the images of hatching eggs between 400 and 1000 nm. The characteristic information of the ratios of length to short axis, the elongation, the roundness and the ratios of the yolk area to the whole area was extracted based on the images. Based on the comparison of the calibration results among 3 waveband regions (400-760, 760-1000, and 400-1000 nm), the visible light in band range of 400-760 nm was chosen to classify actual type of hatching eggs. Different spectra pretreatment methods were used to analyze the spectra, e.g. multiplicative scatter correction (MSC), normalize, standard normal variate transformation (SNV), first derivative (FD), MSC+FD, SNV+FD, normalize+FD, among which the normalized pretreatment method was the most effective, and its classification accuracy was better than other methods. The normalization method was used as the spectral data preprocessing, and then 155 spectral characteristic variables were extracted from 520 wavebands through the correlation coefficient method. Principal component analysis (PCA) method was adopted to reduce the dimension of image-spectrum fusion information which consisted of 4 image characteristic variables and 155 spectral characteristic variables, and then the top 6 principal components were extracted. According to the distribution principle of 2:1 for 300 hatching eggs, the numbers of eggs for training set and testing set were 200 and 100 respectively. RVM and support vector machine (SVM) were used to establish classification models, which were based on image, spectrum and image-spectrum fusion information respectively. The accuracies of the RVM models were 90%, 91% and 96% respectively, while the accuracies of the SVM models were 84%, 90% and 93% respectively. The cost time of the RVM models was 0.6619, 1.0821 and 0.5016 s respectively, while that the SVM models was 5.9386, 5.9886 and 5.6672 s respectively. The experimental results showed that the model based on image-spectrum fusion information was better than the single information model; the RVM model was superior to the SVM model for detecting fertile information of hatching eggs before incubation; and the cost time of RVM model was shorter than that of SVM model. The fertile and infertile eggs were identified very quickly. This project implementing would provide theoretical basis for the real-time online detection and testing of hating eggs for the instrument. Thus using hyperspectral fusion information and RVM can improve the detection accuracy of hatching eggs before incubation.%为了尽可能早的检测出无精蛋和受精蛋，该文提出采用透射高光谱成像技术，融合图像和光谱信息，对其受精信息进行检测。利用高光谱图像系统采集孵化前种蛋在400～1000 nm的高光谱图像，提取图像特征（长短轴之比、伸长度、圆度、蛋黄面积与整蛋面积之比）；筛选出400～760 nm的波段，通过Normalize预处理结合相关系数法提取155个光谱特征变量；运用主成分分析法对图像和光谱的融合信息进行降维，采用相关向量机（relevance vector machine，RVM）分别建立基于图像、光谱和图像-光谱融合信息的受精蛋和无精蛋分类判别模型，并与支持向量机（support vector machine， SVM）模型进行比较，RVM模型检测正确率分别为90%，91%，96%；测试集检测时间分别为0.6619，1.0821，0.5016 s。SVM模型检测正确率分别为84%，90%，93%；测试集检测时间分别为5.9386，5.9886，5.6672 s。结果表明，基于图像-光谱融合所建立的模型优于单一信息的模型，在分类精度上，采用RVM分类精度高于SVM的分类精度；在分类时间上， RVM的分类时间比SVM短，因此，利用高光谱融合信息和相关向量机可以提高种蛋检测精度，研究结果为孵前无精蛋和受精蛋的在线实时检测提供参考。
展开▼
