Probability Graph Model Based on Depth Feature Extraction and Scene Feature Mining and its Application in Advanced Manufacturing

ABSTRACT

In today's advanced manufacturing environment, it is increasingly important to ensure the quality of manufactured products through accurate, automated defect detection. Conventional computer vision and deep learning models (i.e., CNN and VGG19) mainly use the texture of 2D images, often overlooking the geometric depth and contextual relationships necessary for accurate defect detection. To address these limitations, we propose a Probability Graph Model built on Depth Feature Extraction and Scene Feature Mining. The proposed framework utilizes a modified ResNet‐50 backbone to extract multilevel geometric and structural features from RGB‐D images while also employing a Graph Attention Network (GAT) to capture spatial dependencies and contextual scene information from the image regions across the data. Depth and scene features are integrated into a probabilistic model using belief propagation to obtain joint probabilities, allowing for reliably classification. The suggested technique is validated by using the Kaggle Casting Defect Dataset, and it has been demonstrated that our system surpasses the existing deep learning algorithms or methods of CNN and VGG19. The results we obtained indicate that the accuracy of the performance reached 98.54%, precision 98.99%, recall 98.35%, F1‐score 98.43%, and AUC performance 99.13% for intelligent decision‐making. The depth‐aware and context‐aware reasoning investing brings in increased inferences, robustness, and reliability in the industrial setting. This study provides an intelligent framework for real time quality inspection and thus,