Efficient vision defect detection model for drill pipe thread

As an important part of the connection between drill pipes, the thread is easily damaged due to the complex stress and corrosive environment during petroleum exploration and production. Timely and accurate detection of various thread defects is a crucial guarantee for the production safety of oilfield enterprises. However, the widely used manual visual inspection methods and nondestructive detection methods have low efficiency and poor precision in long-term industrial applications. Thus, to solve the difficulty of thread defect detection, we propose an efficient vision defect detection (EVDD) model based on convolutional neural networks to accurately locate and classify the thread defects. The proposed EVDD model is composed of three stages: image enhancement, feature extraction and fusion, and defect prediction. In the first stage, we design a dynamic range image enhancement paradigm to improve the image quality by weighted synthesis of multi-image, which facilitates the feature extraction of the subsequent stage. In the second stage, we construct a convolutional vision transformer module to globally model the spatial dependence among pixels in the feature map through the self-attention mechanism, which provides higher-quality features for the subsequent prediction stage. In the third stage, we propose a multi-branch decoupling prediction head to predict the position, confidence, and category of the defect target separately using three decoupling branches, which improves the information expression ability of the prediction head. Besides, we propose an effective bounding box regression loss function, allowing our EVDD model to learn the morphology-varying rule of thread defects effectively in the training. Extensive experiments show that the proposed EVDD model achieves excellent performance for thread defect detection and better convergence in training iterations. We achieve the best AP (%) of 59.4 on the defect dataset, confirming its great potential to be a general defect detection framework.