Particleboard Surface Defect Inspection Based on Data Augmentation and Attention Mechanisms

Inspection accuracy of surface defect is very important for particleboard production. However, the insufficient defect samples seriously restrict the quality of vision and deep learning-based inspection result. The small-scale defects on particleboard surface are also a major challenge to the input of network models. This paper proposes a method based on data augmentation and attention mechanisms to solve these problems. A hardware platform was designed to take surface defect images. The methods of traditional data augmentation and GAN have been applied to increase the amount of defect samples. The Poisson Fusion technique was adopted to generate defect images albeit varied backgrounds to for network training. The SSD network was deployed as the optimization model. The devised optimization schemes replaced the feature extraction network (VGG) with ResNET18 and ResNET50 respectively before fusing with the DCGAN module. During the training stage, a transfer learning-based method was developed to pre-train the optimized network through COCO2017 dataset to improve the training speed and accuracy. The experimental results showed that the scheme of "ResNET50 + Attention"outperformed benchmarked solutions with a peak performance on particleboard surface defect inspection reaching 96.79%.