MSB R-CNN: A Multi-Stage Balanced Defect Detection Network

Deep learning networks are applied for defect detection, among which Cascade R-CNN is a multi-stage object detection network and state of the art in terms accuracy efficiency. However, it still challenge to deal with complex diverse defects, as widely varied shapes defects lead inefficiency traditional convolution filter extract features. Additionally, imbalance features, losses samples cause lower accuracy. To address above challenges, this paper proposes balanced (MSB R-CNN) based on R-CNN. Firstly, deformable adopted different stages backbone improve its adaptability varying defect. Then, features obtained by refined enhanced feature pyramid. overcome classification regression loss, L1 loss at correct it. Finally, sample selection, interaction union (IoU) sampler online hard example mining (OHEM) combined make sampling more reasonable, can bring better convergence effect model. The results our experiments DAGM2007 dataset has shown that achieve mean average precision (mAP) 67.5%, an increase 1.5% mAP, compared