基于种蛋图像血线特征和深度置信网络的早期鸡胚雌雄识别

为了实现孵化早期鸡胚雌雄识别,构建了机器视觉采集系统,在LED光源下获取180枚鸡种蛋孵化第4天的图像。首先对鸡种蛋图像进行RGB分量提取、中值滤波、感兴趣区域提取等预处理,然后利用限制对比度自适应直方图均衡化、形态学处理、最大类间方差阈值分割和八连通域去噪等方法凸显血线纹理,并通过方向梯度直方图(histogram of oriented gradient,HOG)提取图像的全信息特征和利用灰度共生矩阵提取能量、对比度、相关性、熵、均匀度等5个特征,对HOG全信息特征采用主成分分析(principal component analysis,PCA)降维,最后利用全信息特征和PCA降维特征-灰度共生矩阵特征组合的简化特征,分别构建支持向量机(support vector machine,SVM)、反向传递(back propagation,BP)神经网络、深度置信网络(deep belief networks,DBN)3种鸡胚雌雄识别模型,并比较不同模型的识别准确率。试验中,全信息特征比简化特征构建的模型识别准确率高,基于简化特征的BP、SVM、DBN模型测试集识别综合准确率分别为51.67%、60%和58.33%,基于全信息特征的BP、SVM、DBN模型测试集识别综合准确率分别为58.33%、63.33%和83.33%。其中,基于全信息特征的DBN模型识别准确率最高,达到83.33%。结果表明机器视觉技术为孵化早期鸡胚雌雄识别提供了一种可行方法。

Female and male identification of early chicken embryo based on blood line features of hatching egg image and deep belief networks

Abstract: Nondestructive testing is usually carried out in two directions from the visual features and intrinsic mechanism features of the object. In order to identify the early embryonic sex of chicken eggs, a machine vision image acquisition system was constructed in this study. Under the light source of LED, the development pattern of the egg embryo in different incubation period was collected, and the images of 180 chicken eggs were obtained in 3th, 4th, 5th, 6th, 8th and 10th day. According to the principle of the definition of blood line and blood line integrity in the field of vision of the machine vision, the image of 4th day of incubation was determined to be used to identify the male and female of the embryo. The preprocessing of chicken embryo egg image was carried out, such as component extraction, median filtering, and region of interest (ROI) extraction, followed by the use of contrast limited adaptive histogram equalization (CLAHE), morphological processing, threshold segmentation of Otsu and eight connected domain denoise method to highlight the blood line texture. Through the gray level co-occurrence matrix to extract 5 dimensional features and the direction of the gradient histogram (HOG) to extract 2916 dimensional full information image features. In order to reduce the computational complexity, the processed image was sampled, compressed to a size of 35 pixels × 35 pixels, and the full information features of the 1225 dimension were extracted. Finally, the simplified features of 96 dimensional which was combination of PCA dimensionality reduction--gray level co-occurrence matrix were used to construct three types of chicken embryo eggs male and female identification model which were support vector machine (SVM), 3-layer back-propagation (BP) neural network, 4-layer the deep belief networks (DBN). Also the full information features of 1225 dimensional which was combination of PCA dimensionality reduction--gray level co-occurrence matrix were used to construct three types of chicken embryo eggs male and female identification model which were support vector machine (SVM), 3-layer back-propagation (BP) neural network, 7-layer the deep belief networks (DBN). In the experiment, the information feature of the image was more accurate than that of the simplified feature in the same model, and the recognition accuracy of the whole information feature - DBN model was the highest, reaching 83.33%. Among them, the accuracy of male identification was 76.67% and that of female identification was 90%. The discrimination time of the three models was analyzed for the test set samples, the discriminant time of the three models was SVM, BP, DBN in order of shortest to longest. Correspondingly, the higher the dimension of the input features, the longer the discriminant time of the model, and finally with the highest recognition accuracy which was the full information feature -- the DBN model had the longest discriminant time, which was 7.8350s. The results showed that the machine vision technology provided a feasible method for sex determination of early hatching of chicken embryo eggs.