Sericulture is the process of cultivating silkworm cocoons for the production of silks. The quality silk production requires quality seed production which in turn requires accurate classification of male and female pupa in grainage centers. The challenges in the current methods of silkworm cocoon sex classification using manual observation lie in the time-consuming nature of the process, potential human error, and difficulties in accurately discerning subtle morphological differences between male and female cocoons. FC1 and FC2 single hybrid variety breed pupa are commonly used in south India for the production of high yielding double hybrid bivoltine silkworm seeds. In this study, 1579 FC1 and 1669 FC2 variety samples were used for the classification process. To overcome the challenges of present physical observation by expert employees, camera images of FC1 and FC2 cocoons were used in this study for sex classification. The proposed model used Histogram Oriented Gradient (HOG) feature descriptor of cocoon samples. Linear Discriminant Analysis (LDA) was applied on the feature vector to reduce the dimension and this feature matrix was given to the classical machine learning algorithms support vector machine (SVM), k-nearest neighbors (kNN), and gaussian naïve bayes for classification with stratified 10-fold cross validation. The results showed that for FC1 data HOG + LDA + Naïve Bayes performed better with a mean accuracy of 95.3% and for FC2 data HOG + LDA + KNN attained a mean accuracy of 96.2%. Our results suggest that this camera imaging method can be used efficiently in the classification based on the cocoon size and shape of different breeds.

养蚕是培育蚕茧以生产丝绸的过程。高质量的丝绸生产需要高质量的种子生产，这反过来又需要谷物中心对雄性和雌性蛹进行准确的分类。目前通过人工观察进行蚕茧性别分类的方法面临的挑战在于过程耗时、潜在的人为错误以及难以准确辨别雄性和雌性茧之间细微的形态差异。 FC1和FC2单杂交品种蛹在印度南部普遍用于生产高产双杂交二性蚕种子。在本研究中，1579 个 FC1 和 1669 个 FC2 品种样本用于分类过程。为了克服目前专家员工物理观察的挑战，本研究使用 FC1 和 FC2 茧的相机图像进行性别分类。所提出的模型使用茧样本的直方图定向梯度（HOG）特征描述符。对特征向量应用线性判别分析 (LDA) 以减少维度，并将该特征矩阵提供给经典机器学习算法支持向量机 (SVM)、k 最近邻 (kNN) 和高斯朴素贝叶斯进行分类分层10倍交叉验证。结果表明，对于 FC1 数据，HOG + LDA + Naïve Bayes 表现更好，平均准确度为 95.3%；对于 FC2 数据，HOG + LDA + KNN 平均准确度为 96.2%。我们的结果表明，这种相机成像方法可以有效地用于根据不同品种的茧大小和形状进行分类。