Infrared Image Denoising Techinic Research Based On Convolutional Neural Network

Compared with the optical image,the infrared image suffered from dim targets background clutter.The utilization of infrared noisy images will adversely impact the subsequent image processing,especially when carrying out feature learning of images via deep learning technology.Misdiagnosis might occur in both image recognition and image classification.When low fault tolerance is required,such as in military and medical fields,the impact of these noises can be extremely fatal.Therefore,it is significant to eliminate such noise of infrared image to improve the image quality.Currently,the mainstream in infrared image denoising algorithms tends to eliminate the edge information of the image as well when removing noise,damaging the original features of the infrared image and setting more obstacles to the corresponding image processing later.Among existing methods,image feature can be learned automatically by image denoising methods based on convolutional neural network.This method can extract the information in image characteristics and learn quickly after training,so as to complete infrared image denoising fast and efficiently.It plays a prominent role in protecting the details of infrared image information.In this thesis,a method of infrared image denoising based on multi-scale parallel convolution neural network is proposed,which uses multi-scale convolution and multi-scale deconvolution for upsampling and downsampling processes,and consists of a residual network section and a dense network section in parallel in the feature sampling section.The residual network is composed of bottleneck residual blocks and the dense network is made up of dense blocks.Experiments show that the proposed multiscale parallel convolutional neural network in this research has higher Peak Signal to Noise Ratio and Structural Similarity Index Measurement than other convolutional neural network methods,and more clear in visual.