BV-Net: Bin-based Vector-predicted Network for tubular solder joint detection

Tubular solder joint detection is an important and challengeable issue in industry, due to the small objects, rarely collected datasets and real-time and high precision requirements. Traditional methods on defect detection cannot solve tubular solder joint detection due to lacking of angle estimation. In this paper, we propose a tubular solder joint detection method named Bin-based Vector-predicted Network (BV-Net), which combines the framework of state-of-the-art deep-learning-based object detector (YOLOv4) with specific characteristics and requirements of tubular solder joint detection. BV-Net could effectively estimate both the center point and the direction of tubular solder joints. Firstly, To regress the center point, we propose a Circle-based Distance-Intersection over Union (CirDIoU) loss, which gets better learning performance for the center point of tubular solder joint than Distance-Intersection over Union (DIOU) loss. Secondly, to estimate the direction, we introduce a bin-based angle regression method, which transforms a regression task into a classification and regression task, improving the precision of direction estimation greatly. Thirdly, we establish a tubular solder joint dataset and design a new evaluation index: mAP (δd, δθ) for tubular solder joint detection, combining the relative deviation of center point positioning δd and the relative deviation of angle regression δθ. Finally, comparison experiments on the dataset are carried out. BV-Net achieved 85.5% mAP (0.5%, 3%) with 34.4 FPS, meeting the requirements of industrial system. In direction estimation, bin-based angle regression method promotes 4.3% mAP (-, 3%), compared with the baseline. In center point positioning, BV-Net outperforms YOLOv4 by an improvement of 0.4% mAP (0.5%, -). The experimental results verified the effectiveness of our method.