Unleashing the power of deep learning: A novel approach for defect detection in solid rocket motor

High-Energy Materials (HEMs), particularly those integral to solid rocket motors are linchpins of mission success and safety. Detecting defects within a solid rocket motor is challenging but crucial for performance in aerospace and defense. Traditional methods and human inspection are inadequate, prompting the need for innovative defect detection. Our study introduces a deep learning approach, departing from conventional methods. Artificial intelligence-based automation is crucial for addressing the complexities and limitations of traditional defect detection methods. However, existing methods still grapple with challenges, in cases where images are non-defective, the conventional approach involves unnecessary annotation, leading to laborious efforts and computational costs. In some scenarios, low-quality images compromise feature extraction, impacting detection accuracy. To address this, our study introduces a novel approach. We employ a two-step process image classification and subsequent defect detection enhanced by dedicated neural network models. We start by classifying images as defective or non-defective using a unique combination for image classification that leverages the power of the Convolutional Autoencoder (CAE) in conjunction with the Visual Geometry Group Network-19 (VGG-19) architecture (CAE+VGG19). Only when an image is classified as defective, it proceeds to object detection and segmentation by Detectron2. Non-defective images bypass this step, thus saving valuable time and computational costs. This approach, not commonly found in previous studies. Our methodology achieves a remarkable 98% accuracy in image classification, in defect detection and segmentation, by Detectron2, accuracy exceeds 90%, ensuring precise localization. This study represents a significant advancement in Solid rocket motors (SRMs) defect detection, providing a precision-driven, efficient, and cost-effective solution.