This technology downsamples input images into various sizes, applies super-resolution restoration algorithms—including TSC, SCDL, and CDL sparse coding, as well as SRCNN and VDSR deep learning—and then performs LBP or VGG facial recognition on each restored image to automatically select the algorithm, recognition technique, and upscaling factor that yield the highest accuracy.
Interpolation-based super-resolution is computationally simple but results in blurry images; reconstruction-based methods cause stair-step artifacts along edges; and learning-based methods require massive training datasets. Furthermore, there has been a lack of criteria for determining which restoration technique is most effective for facial identification in low-resolution surveillance footage.
By comparing recognition results across combinations of 10×10, 15×15, and 20×20 subsampling and 2x, 3x, and 4x upscaling, this technology identifies the optimal settings, ensuring restored images closely resemble the original and achieving facial recognition accuracy near that of the source. It can be applied to security CCTV, access control, and public safety monitoring systems to restore distant or low-quality footage to a level suitable for identification.
N/A
