Denoising watermarked images using Bidirectional Recurrent Convolutional Neural Networks
Pragya Agarwal, Somesh Kumar, Chandra Shekhar Yadav · Research Square · 2023
Abstract Recurrent Neural Networks are being widely used for examining and learning sequential motion as well as stationary data, like video, audio, speech, images, etc. In this paper, a Bidirectional convolutional Recurrent Neural Network is being used, for the purpose of image denoising for watermarked images. A Bidirectional network has two parallel channels and can store and memorize information from both forward and backward directions. Because of the information taken from an extended time zone, it can learn and predict more accurate results and can model better visual-temporal dependency. Moreover, instead of considering a fully connected recurrent neural network from input to hidden layer, a weight-sharing convolutional neural network is employed, which is less costly and that’s why the model can become more computationally efficient. The input for this denoising technique is a watermarked image, which is overall the greatest challenge. There are many other existing image and video denoising methods, based on Convolutional Neural Networks, Recurrent Neural Networks, and some other traditional methods. But none of them is able to denoise a watermarked image/video effectively, because they consider the inserted watermark as a sequence of noise and try to destroy it. In this paper, the performances of RNN and CNN based techniques are analyzed and compared with the proposed image denoising technique, which is based on BRCNN. It has been observed experimentally that the proposed technique’s accuracy rate is much more (95 percent) than its counterpart techniques RNN and CNN (89 percent and 67 percent respectively). Moreover, the time taken for removing noise is less (9.35 Seconds) as compared to RNN based techniques (15.29 seconds) and CNN-based techniques (12.67 seconds).