A Text Image Super-Resolution Generation Network without Pre-training

Aiming at the problem of low recognition rate of optical character recognition for low-resolution text images, a non-training text image super-resolution generation network is designed. The network is mainly composed of a convolution layer, a Batch Normalization layer (BN layer), a LeakyReLU activiation layers, an upsampling layer, and a downsampling layer. The network does not need to use the dataset for training. It can transform the super-resolution reconstruction problem into an optimization problem and process the image directly. A new loss function for text image super-resolution to realize super-resolution reconstruction of text image is proposesing. The experimental results show that the network can achieve 73.30% OCR accuracy on the ICDAR 2015 TextSR dataset, which is 12.66% higher than the bicubic algorithm. The network also has better accuracy to the learning-based Adaptive Sparse Representation Selection algorithm.