To train the VDSR network, set the doTraining variable in the following code to true. The pretrained network enables you to perform super-resolution of test images without waiting for training to complete. īy default, the example loads a pretrained version of the VDSR network that has been trained to super-resolve images for scale factors 2, 3 and 4. Specify dataDir as the desired location of the data. This function is attached to the example as a supporting file. Use the helper function, downloadIAPRTC12Data, to download the data. Then, go directly to the Perform Single Image Super-Resolution Using VDSR Network section in this example. If you do not want to download the training data set, then you can load the pretrained VDSR network by typing load("trainedVDSRNet.mat") at the command line. The data set includes photos of people, animals, cities, and more. Download Training and Test Dataĭownload the IAPR TC-12 Benchmark, which consists of 20,000 still natural images. Additionally, the VDSR network can generalize to accept images with noninteger scale factors. Scale augmentation improves the results at larger scale factors because the network can take advantage of the image context from smaller scale factors. This example trains a VDSR network with multiple scale factors using scale augmentation. VDSR solves this problem by using a large receptive field. As the scale factor increases, SISR becomes more ill-posed because the low-resolution image loses more information about the high-frequency image content. If Y high res is the luminance of the high-resolution image and Y lowres is the luminance a low-resolution image that has been upscaled using bicubic interpolation, then the input to the VDSR network is Y lowres and the network learns to predict Y residual = Y highres - Y lowres from the training data.Īfter the VDSR network learns to estimate the residual image, you can reconstruct high-resolution images by adding the estimated residual image to the upsampled low-resolution image, then converting the image back to the RGB color space.Ī scale factor relates the size of the reference image to the size of the low-resolution image. VDSR is trained using only the luminance channel because human perception is more sensitive to changes in brightness than to changes in color. In contrast, the two chrominance channels of an image, Cb and Cr, are different linear combinations of the red, green, and blue pixel values that represent color-difference information. The luminance channel of an image, Y, represents the brightness of each pixel through a linear combination of the red, green, and blue pixel values. The VDSR network detects the residual image from the luminance of a color image. A residual image contains information about the high-frequency details of an image. In the context of super-resolution, a residual image is the difference between a high-resolution reference image and a low-resolution image that has been upscaled using bicubic interpolation to match the size of the reference image. VDSR employs a residual learning strategy, meaning that the network learns to estimate a residual image. This mapping is possible because low-resolution and high-resolution images have similar image content and differ primarily in high-frequency details. The VDSR network learns the mapping between low- and high-resolution images. VDSR is a convolutional neural network architecture designed to perform single image super-resolution.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |