Image compression algorithm research based on improved LSTM
Shasha Li, Mengjun Li, Pengfei Li, Yongjun Li · 2020
With the advent of the era of big data, storing and transferring data is facing tremendous pressure. How to use deep learning to obtain higher compression ratio and higher quality images has become an urgent problem. Recurrent neural network (RNN) can control the bit rate of images with iterative manner to improve compression performance. However, RNN needs to introduce long short term memory (LSTM) to solve the problem of long-term dependence, which leads to the model more complex. In order to speed up training process and reconstruct higher-quality images, firstly, this paper improves the activation function in LSTM to better determine the information to be stored or forgotten, so that the amount of parameters is reduced and the training process is faster. Then, the image recovery block is introduced in the decoder to reconstruct high-resolution images. Finally, instead of L1 loss, we use SmoothL1 loss to accelerate the convergence of loss. Experimental results show that our model can achieve a higher compression ratio, and evaluated by SSIM the value is more nearly to 1.