An overview of IndSR dataset composed by pixel-aligned LR-HR image pairs which includes five common defects (pit, scratch, sand hole, shortage and mixture) with pixel-level mask in industrial castings

    The acquisition and processing of visual data in industrial environments are of paramount importance. High-resolution images (HR) offer superior clarity and richer textural detail when compared to low-resolution images (LR). On the one hand, HR images can be utilized to enhance downstream tasks such as defect detection, target segmentation, and target detection. On the other hand, they provide invaluable insights to designers and quality inspectors who require a detailed understanding of the images. Currently, the majority of research on super-resolution algorithms focuses on natural scenes such as cities and fields, however, the development of datasets for industrial scenes is still in its infancy. To address significant variations in the reflectivity of the surfaces and image distortion in building real HR-LR images, we design a capture to registration framework for realistic image super-resolution in the industry environment. It consists of the standard imaging system, physical calibration of the imaging system, as well as the rigid and elastic registration of the LR and HR images. Thus, we build the first real industrial super-resolution dataset (IndSR), comprises of 200 sets of calibrated images with three scaling factors and with different defects. To benchmark the IndSR, we employ quantitative, qualitative, and task-oriented studies to evaluate the representative super-resolution algorithms and defect detection methods. Besides, we systematically investigate and discuss the performances and results of the existing methods to the potential to advance research in the field of super-resolution algorithms in industrial scenes.

Overview of Capture-to-Registration Framework and benchmark: The purposed realistic image capture system captures images at multiple scale factors by adjusting the imaging distance. We obtain aligned LR image pairs through rigid to elastic registration with HR images serving as ground truth. To conduct benchmark testing, we first qualitatively assess the performance of various algorithms trained on simulated and real datasets. Subsequently, we conduct quantitative evaluations of super-resolution effects using both full-reference and no-reference measurements. Finally, we undertake an evaluation task that focuses on defect detection.

×4

×3

×2

HR

Mask

Average full-reference and no-reference metrics on our IndSR testing set by different SISR methods(trained on the synthetic dataset and IndSR dataset respectively).

SR results (×4) on our IndSR testing set by different SISR methods (trained on the synthetic dataset and IndSR dataset respectively).

The test results of the origin and super-resolution image (different methods trained on synthetic and realitic IndSR datasets) in the anomaly detection network.