Data Augmentation and Layered Deformable Mask R-CNN-Based Detection of Wood Defects

The detection of wood defects plays an important role in the processing and production of wood, the deep learning methods have achieved outstanding results in this field in recent years. There are still two unresolved problems, one is that the category of defects is unbalanced and difficult to obtain, and the other is the effective modeling of defects with different sizes and irregular shapes. Here an alternative detection method is proposed. For the problem of imbalance in defect categories, a generative data augmentation method, cycle generative adversarial network (GAN), is used to generate new defective images, while seven online data augmentation schemes are used to solve the problem of difficult access to defective data. The layered deformable mask region-based convolutional neural network (Mask R-CNN) is constructed to address the challenges of modeling irregular defects and contextual information extraction. By creating layered connections between residual modules, a larger receptive field is achieved at each layer of the network, while deformable convolution is used to better fit the shape of the defects. Finally, the upgraded model is trained and used to locate and segment defects simultaneously. Ablation experiments show that the enhanced model performs better than the original model in terms of localization and segmentation. Meanwhile, it is excellent relative to baseline object detection algorithms, instance segmentation algorithms, and other methods for detecting wood defects.

Automated summary

In this study, a novel detection method using generative adversarial networks and data augmentation is proposed to address the issues of imbalance in defect categories and difficult access to defective data in wood defect detection. Additionally, a layered deformable mask region-based convolutional neural network is constructed to tackle the challenges of modeling irregular defects and extracting contextual information.