3.1. Data Collection and Pre-Processing

The particleboard images were collected from the particleboard production facility of Fenglin Yachuang Group in Huizhou, China. The original picture size of the five different types of defective particleboard was 1751 by 911 pixels, with a bitmap depth of 8 bits, for a total of 982 images. These photos are split into three groups in a ratio of 7:2:1, and seven image modifications, including mirroring, grayscale conversion, median filtering, white noise, poisson noise, gaussian noise and pretzel noise are applied to each group. The dataset has 7856 total photos and is made up of both the enhanced data and the original photographs. Among them, 70% of the training set images are 982 × 0.7 × 8 ≈ 5499, 20% of the validation set images are 982 × 0.2 × 8 ≈ 1571 and the remaining 10% are used as the testing set images, which are 982 × 0.1 × 8 ≈ 786. This is shown in Figure 5.

Data augmentation and annotatione.

Two programs, LabelImg and LabelMe, are used to generate target detection labels and semantic segmentation labels for five different types of defects seen on the surface of particleboard, including SandLeakage, BigShavings, GlueSpot, OilPollution, and Soft. Each area of interest is covered by the label, which also pinpoints the defective pixels. Table 1 provides details on the surface flaws in particleboard.

Specific information of particleboard surface defect dataset.