Design And Implementation Of Image Enhancement System Based On Deep Learning

Image enhancement is a research hotspot in the field of image processing.The main task is to process images that affect human perception.Through image enhancement,the image can restore more information,and the image quality can also be improved.Video surveillance,as the most common application in daily life,will be affected by various adverse factors such as harsh environment,electromagnetic interference,etc.,which will lead to the problems of unclear,noisy and low video resolution of the surveillance video.The low-quality surveillance image will affect the surveillance performance,so it becomes particularly important to enhance the surveillance image.Based on the generative adversarial network,this thesis integrates two tasks,image noise reduction and image super-resolution.In this system,image noise reduction is achieved by a channel attention mechanism.For the image resolution part,we can provide different upsampling rates to reduce bandwidth pressure in surveillance video transmission.Finally,combining image noise detection algorithms and image enhancement models,an image enhancement system that can be applied to video surveillance is given.The main research work of this thesis is as follows:(1)A noise detection algorithm suitable for this image enhancement system was proposed.Four directions of Laplacian were used to convolve with the original image.The difference between the calculation result and the original image pixels is used to determine whether it is a noise point.(2)An image enhancement model capable of image noise reduction and super resolution was proposed.The entire model was trained in an adversarial training mode to obtain better results.The model includes three modules:feature extraction,feature noise reduction and image reconstruction.The feature extraction module obtains receptive fields of different sizes by using the structure of short jump connections.The feature noise reduction module is used to help the network to identify noise and achieve noise reduction,the key property of which is that we introduced a channel attention mechanism into the module.The image reconstruction module performs super-resolution of the image by using sub-pixel convolution.The experimental results show that the model can effectively learn the noise distribution of the image and restore more texture details of the image.(3)An image enhancement prototype system was designed and implemented.The system mainly includes two major functional modules,noise detection and image enhancement.The results of the system performance test show that the system proposed in this thesis can identify image noise and perform image enhancement according to customer needs,and has good engineering feasibility.In summary,the main work of this thesis is to study an image enhancement model based on generative adversarial networks,which can improve image quality by carrying out simultaneous noise reduction and super-solution.Finally,an image enhancement system was designed and implemented.