Nowadays, Optical Character Recognition (OCR) can be used to translate text image into digital text instead of human effort. However, noises in text images, such as strikeout or faded-out text, can make output from OCR technology contains wrong content. This project aims to make OCR technology generate text accurately by denoising strike-out in text image. We provided 4 models with 2 architectures: Convolution with Max Pooling, and Convolution and Deconvolution Autoencoder. The result showed that Convolution with Max Pooling together with residual net had the best performance with minimum loss of 0.0016%.