A performance and quality evaluation of residual learning deep convolutional neural networks for image denoising applications based on the architecture outlined in the paper IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 26, NO. 7, JULY 2017

• 50x50x1 input layer, the input of our DnCNN is a noisy observation modeled by y = x+v
• 20 Convolution, ReLU, and normalization layers:
  • 2d convolution w/ 64 3x3x64 convolutions, weights/bias initialized to zeros, and uniform padding across all dimensions, stride set to 1
  • Rectified Linear Unit activation layer (ReLU)
  • Batch normalization layer
• MSE regression output Layer
• Intuition:
  • Theoretically increasing the convolution layers further generalizes features, thus increasing ability to delineate from noise

• 75 test images, where each image was degraded via Gaussian noise with variance randomly sampled between 0.005 and 0.2
• Objective performance metrics used:
  • SSIM: Correlated with quality/perception of human vision
  • PSNR: Evaluates noise based on mean square error to test image
  • Benchmark filters used as comparison:
    • Median, Gaussian, Wiener, Wavelet, Non-local means, bilateral filter

• Image Processing Toolbox
• Deep Learning Toolbox
• Parallel Compute Toolbox