Research On Image Enhancement Algorithm Using Convolutional Auto-Encoder And Its Parallel Implementation

With the rapid development of the Internet,information has exploded.In the face of massive data sets,how to use existing data,find useful information for people and maximize the potential value of data has become a widespread concern in research and industry.Traditional image enhancement algorithms have been unable to meet the needs of massive data sets.People need new image processing frameworks,and deep learning provides a new direction for image enhancement algorithms.In this paper,the image enhancement method LLNet based on self-encoder is studied and improved.Aiming at the shortage of a large number of redundant parameters when applying to multi-channel color images,an image enhancement method based on convolutional self-encoder is proposed.-CAENet,this method is well suited for three-channel color images and has strong robustness.Then,in order to apply the neural network algorithm CAENet to the big data background,this paper then proposes a distributed framework to implement the algorithm.At the same time,in order to reduce the parameter fluctuations generated during neural network training,the network will reach a faster and more stable convergence.In this paper,based on the exponential moving average algorithm,an optimization algorithm called truncated exponential moving average is proposed.Finally,the image enhancement method CAENet is applied to the proposed parallel framework,and finally satisfactory results are obtained.The main work of this paper is as follows:(1)For the neural network-based image enhancement method LLNet(Low light network)will generate a large number of redundant parameters when it is extended to the three-channel color image in the actual scene.This paper proposes a convolutional self-encoder based on convolution.Neural network image enhancement method-CAENet(Convolutional Auto-Encoder Network).Experiments prove that CAENet can effectively improve the image light perception and color perception while retaining the image details,and has a good effect in overcoming the shortcomings of excessive color saturation and uneven color patches.In addition,for the noisy low-light image,the image can be enhanced while achieving the denoising effect,showing that CAENet has strong robustness.(2)In order to apply the neural network algorithm well in the context of big data,this paper proposes an optimization algorithm named Trim Ecponential Moving Average,which can smooth the network training well.At the same time,in order to reduce the time of network training,this paper combines neural network with distributed mode,then proposes a neural network parallel framework based on TEMA.The network is well distributed and speeds up the training of neural networks.Finally,the experiment proves that the framework can be well applied to neural network training calculation.(3)Applying the image enhancement method CAENet to the distributed framework proposed earlier in this paper,the necessity of image enhancement application in distributed is expounded.The final experiment proved that CAENet saves a lot of time during training,which provides an effective solution for image enhancement applications in the context of big data.(4)Finalize the full text and propose a prospect for further research.