Identification of Green Gram (Vigna radiata) Grains Infested by Callosobruchus maculatus Through X-ray Imaging and GAN-Based Image Augmentation

• Published in: Journal of Biosystems Engineering Vol. 47; no. 3; pp. 302 - 317
• Main Authors: Divyanth, L.G., Chelladurai, V., Loganathan, M., Jayas, Digvir S., Soni, Peeyush
• Format: Journal Article
• Language: English
• Published: Singapore Springer Nature Singapore 01.09.2022; 한국농업기계학회
• Subjects: Agriculture; Control; Engineering; Environmental Engineering/Biotechnology; Mechatronics; Original Article; Robotics; 농학; Green gram; Support vector machine; Grain quality; Convolutional neural network; X-ray imaging
• ISSN: 1738-1266; 2234-1862
• DOI: 10.1007/s42853-022-00147-9

Purpose Green gram ( Vigna radiata ) is an important food legume of the world. However, post-harvest losses due to pulse beetle, Callosobruchus maculatus (F.), are significant due to improper storage management practices and undetected internal infestations. The detection of early stages of infestation could help in implementing suitable control practices for insect disinfestation. This study determined the potential of detecting internal infestations caused by C. maculatus using the soft X-ray method and deep learning. Furthermore, this study aims to reduce the time and effort needed to prepare a huge amount of image data for this highly data-driven process by using generative adversarial networks (GANs). Methods A three-class classification method was implemented to identify the infestation stages, namely, uninfested kernel, larva stage, and pupa stage. The approach was based on features extraction from the deepest pooling layer of a state-of-the-art Convolutional Neural Network architecture—the Xception , and using support vector machine as the classifier. Moreover, a GAN model was proposed to synthesize artificial X-ray images. Results The overall F 1-score produced by the model was improved from 0.86 to 0.91 when the GAN-synthesized dataset additionally supported the training data. Also, the classification accuracy for detecting the stage of internal infestation improved by 5.5%. Conclusion The experiment showed that X-ray imaging and deep learning–based automatic features extraction could identify internal infestation in green gram grains. The results determine that augmentation using GANs can enhance the status of learning-based grain quality assessment models with reduced manual effort.