Injection-molded product defect dataset

Due to the lack of publicly available injection-molded product defect datasets and the diversity of defects in terms of shapes, sizes, and textures, we collects defect samples from injection molding factories to ensure the model performs well in real industrial scenarios. To ensure the quality and usability of the data, after analyzing the sample data, data cleaning is performed to remove the irregular images. Considering the redundant backgrounds and high resolutions in the initial images, image segmentation is used to highlight defect features, and the image sizes are standardized to 1280×1280. The defects are then categorized into five classes: flash, spot, scratch, flowingmark, and burning. The number of samples in the scratch, flowingmark, and burning classes is relatively small. To balance the number of samples in different classes, data augmentation methods such as rotation, Gaussian blur, color transformation, and contrast adjustment are applied to the three classes of scratch, flowingmark, and burning to have 500 images in each class. After labeling, a dataset containing 2500 defect images is constructed, and the defect images are randomly divided into training, testing, and validation sets in a 7:2:1 ratio, providing the basis for model training and evaluation.