Image GPT with Super Resolution

AbstractA Generative Pre-trained Transformer (GPT) model which can generate text by looking at previous text was trained to generate image pixels sequentially by making a correlation between the image classification accuracy and the image quality. This model uses the generative model for generating images. The Image Generative Pre-trained Transformer (IGPT) works on a low-resolution image which in turn produces a low-resolution output. In this paper, we have attempted to eliminate this limitation by enhancing the resolution of the output image produced by IGPT. The primary focus during this research work is to check different models and choose the simplest model for improving quality of the image generated because there are several models that support deep neural networks that have been successful in upscaling the image quality with great accuracy for achieving super resolution for a single image. The output image of low resolution is upscaled to high-resolution space employing a single filter and bicubic interpolation. We have also considered peak signal-to-noise ratio (PSNR) score and structural similarity (SSIM) value to analyze the standard of the image produced by the algorithm. The proposed approach has been evaluated using images from publicly available datasets. We have used leaky ReLU instead of ReLU as the activation function which produces better PSNR score and SSIM value, improving the overall result. Combining efficient sub-pixel convolutional neural network (ESPCNN) algorithm with IGPT, we have managed to get better output compared to the output generated by IGPT solely.KeywordsGPTTransformersCNNImage completionImage classificationSuper resolution