Research On Light Field Image Coding Based On Deep Neural Network

With the development of science and technology,the clarity of video data such as video images is getting higher and higher,and they also need more and more storage and transmission resources.In order to ease this burden,we need to do some compression on the image data.The HEVC(High Efficiency Video Coding)standard is one of the image coding methods and it has very good coding efficiency and can compress the image files as much as possible while ensuring the image quality.The HEVC encoding standard encodes an image by dividing the image into blocks,uses the already encoded image blocks to predict adjacent blocks,and records the difference between them to achieve the effect of encoding the image.At present,the use of neural networks for image coding is a hit,and many image coding network models have achieved good results.In this article,the author used the Long Short-Term Memory(LSTM)model and the generated confrontation network(GAN,Generative Adversarial Networks)model to encode light field images.The LSTM model uses a recurrent neural network and it contains three main parts in total,they are called the encoder,the binary network and the decoder.Using a large amount of image data to train the network,a weight parameter for encoding can be obtained.In the encoding process,the encoder will load these weight parameters,and the image will generate an encoding file through the encoder as well as the binary network.During the decoding process,the decoder will also load this weight parameter and input the encoded file into the decoder to get the reconstructed image.The entire encoding process can be optimized through multiple iterations.Bringing each result into the next iteration process can efectively improve the effect of encoding and decoding.The GAN model can learn the characteristics of the input data and generate data based on these characteristics.We input our image data set as training data into the GAN model.The generation network and discriminant network of the GAN model will optimize each other in the process of confrontation.The generating network will simulate and generate similar images according to the characteristics of these images,and the discriminating network will judge the authenticity of the generated data.In this process,the quality of the image generated by the generated network is getting higher and higher.Until the network can no longer determine the authenticity of the generated image,the quality of the generated image is very close to the original image.We use these two methods to encode our image data,analyze the encoding effect of these methods,and make a corresponding summary.