A Machine Learning approach to Silkworm pupae gender identification

---

Abstract

Silkworm production is an important industry worldwide, particularly in regions where silk production is a significant economic activity. The goal is to quickly and accurately identify the gender of silkworm pupae. Visual classification methods of silkworm gender are time-consuming and leading to errors, it reduces productivity. The objective is to develop a software-based solution for gender classification of silkworm pupae using machine learning technique. A machine learning approach of classifying silkworm pupae gender is proposed using a dataset of 950 images. Before getting into transfer learning the dataset was tested with different edge detection techniques like Canny, Sobel and Prewitt to identify the gender based on edges. But it depends on image quality, rotation and reference image. Then the dataset is tried with a feature detection method like Oriented FAST and Rotated BRIEF (ORB) but it is a rotation variant and sensitivity to noise. The dataset is pre-processed and augmented to produce 3000 images in order to improve the accuracy. Conventional Neural Network (CNN) will automatically learn hierarchical features from raw data. Next, a 4:1 split of the dataset is made into training and testing sets. Model training begins with feature extraction using CNN architectures such as MobileNetV2 and ends with the plotting of model evaluation metrics. The accuracy obtained from the MobileNetV2 architecture is 95%.

---

Keywords

Convolutional Neural Network (CNN); Edge detection; Oriented Fast and Rotated BRIEF (ORB); Image augmentation; learning; MobileNetV2; Machine Silkworm pupae; Transfer learning,