Research on low illuminance fluorescence permeation digital image restoration method based on deep learning
Fluorescent penetration testing images are usually collected in extremely low illuminance darkroom environments,which inevitably leads to issues such as image blurring and low contrast.A deep learning based image restoration method is pro-posed.Firstly,a fluorescence penetration detection image dataset is constructed,and a total of 3400 images in Raw format are cap-tured under different short and long exposure times,with 20%being the test set and 10%being the validation set;Secondly,con-struct a convolutional neural network based on U-net,interpolate the Raw format data and divide it into four RGGB channels.Then subtract the black level value and amplify the input data brightness by the corresponding multiple as network input.Finally,the network output is upsampled through sub pixel layers to obtain the output image in RGB space.The test results show that selecting SSIM and MAE as loss functions and using the Adam optimizer for model training,the image PSNR reaches 28.237dB and SSIM reaches 0.783,both of which are superior to traditional methods.It can effectively reduce image noise and improve image clarity,and obtain more missing details in low light images.
万方数据
维普
