An End-to-End Compression Framework Based on Convolutional Neural Networks

• Traditional codecs reduce irrelevance and redundancy by distributing bits among quantized transform coefficients, lowering bits per pixel as compression increases.Source: §I · p. 1
• Iterative nonlocal and low-rank post-processing methods can improve deblocking but are computationally expensive and difficult to apply in practical applications.Source: §I · p. 1, A. Image Deblocking and Artifacts… · p. 2
• The proposed framework combines ComCNN, an image codec, and RecCNN to learn a compact structural representation and reconstruct decoded images end to end.Source: §I · pp. 1–2
• The unified learning algorithm simultaneously trains both CNNs despite the image codec’s nondifferentiable quantization rounding function.Source: §I · pp. 1–2
• Unlike prior methods that focus only on post-processing, the framework jointly optimizes encoder front-end and decoder back-end processing while remaining compatible with JPEG, JPEG2000, and BPG.Source: A. Image Deblocking and Artifacts… · p. 2, §I · p. 2

• CNN-based super-resolution methods use architectures such as SRCNN and benefit from residual learning and gradient-based optimization for training deeper networks.Source: B. Image Super-Resolution Based on… · p. 2
• Deep-learning image compression has addressed both lossy and lossless settings, including variable-rate and arbitrary-size image compression with recurrent neural networks.Source: C. Image Compression Based on Deep… · p. 2
• Prior deep-learning compression methods achieved competitive or state-of-the-art results but generally ignored compatibility with existing image codecs.Source: C. Image Compression Based on Deep… · pp. 2–3
• The proposed framework is illustrated as combining two CNNs with an image codec in an end-to-end compression pipeline.Source: Figure 2 · p. 3, §III · p. 3

A. Architecture of End-to-End Compression Framework

• ComCNN generates a compact representation that preserves image structure for encoding with an image codec.The compact representation is produced from the input image before codec encoding.Source: A. Architecture of End-to-End… · p. 3, 1) Compact Representation… · p. 3
• ComCNN uses three weight layers that maintain spatial structure and support RecCNN-based reconstruction.Its layers perform patch representation, downscaling and feature enhancement, then compact-representation construction.Source: 1) Compact Representation… · p. 3
• RecCNN contains 20 weight layers combining convolution, ReLU, batch normalization, and convolutional output reconstruction.Residual learning and batch normalization are used during training.Source: 2) Reconstruction Convolutional Neural… · p. 3
• RecCNN enhances the decoded image and reconstructs the high-quality output through residual learning.It predicts a residual image that is added to the upscaled decoded input.Source: A. Architecture of End-to-End… · p. 3, 2) Reconstruction Convolutional Neural… · p. 3, Figure 4 · p. 4
• The end-to-end pipeline maps an original image through ComCNN, an image codec, and RecCNN to produce the reconstructed image.The codec may be JPEG, JPEG2000, or BPG.Source: B. Learning Algorithm · pp. 3–4
• The codec’s rounding operation is nondifferentiable, so the framework uses an iterative optimization learning algorithm.This issue arises during backpropagation.Source: B. Learning Algorithm · p. 4

1) Updating the Parameters θ2 of RecCNN:

• An auxiliary variable represents the decoded compact representation of the original image.It is introduced according to the network topology before optimizing RecCNN.Source: 1) Updating the Parameters θ2 of RecCNN · p. 4
• Combining the relevant equations yields the auxiliary-variable update used in the learning procedure.This update follows from combining Eq.(4) and Eq.(3).Source: 1) Updating the Parameters θ2 of RecCNN · p. 4

2) Updating the Parameters θ1 of ComCNN:

• Optimizing ComCNN is difficult because the codec operation is inherently nondifferentiable during backpropagation.The method addresses this difficulty by defining an auxiliary variable.Source: 2) Updating the Parameters θ1 of ComCNN · p. 4
• The auxiliary decoded representation is used as the optimal input of RecCNN in the ComCNN update.The construction relies on a monotonicity assumption concerning RecCNN.Source: 2) Updating the Parameters θ1 of ComCNN · p. 4
• Because the codec is an encoder, the method derives a practical solution for the auxiliary-variable optimization.The derivation combines the codec property with the stated monotonicity assumption.Source: 2) Updating the Parameters θ1 of ComCNN · p. 5
• Equation (13) approximates the original objective and is used to train ComCNN instead of the nondifferentiable formulation.The approximation is obtained under a reasonable assumption and accompanying derivations.Source: 2) Updating the Parameters θ1 of ComCNN · p. 5
• The optimal θ1 and θ2 are obtained by iteratively optimizing their respective objectives in the proposed algorithm.Algorithm 1 initializes the parameters and alternates training of RecCNN and ComCNN.Source: 2) Updating the Parameters θ1 of ComCNN · p. 5

C. Loss Functions

• Mean squared error is used as the loss function for training with original images and trained θ2 parameters.The loss is introduced for the subsequent optimization procedure.Source: C. Loss Functions · p. 5

For ComCNN training:

• The training formulation uses N as the batch size.Source: 1) For ComCNN training · p. 5
• θ1 denotes the trainable parameter in the ComCNN training formulation.Source: 1) For ComCNN training · p. 5
• The passage defines both batch size and trainable parameters for ComCNN training.Source: 1) For ComCNN training · p. 5

For RecCNN training:

• MSE is used as the loss function for compact representations and original images.Source: 2) For RecCNN training · p. 5
• θ2 denotes the trainable parameter in the RecCNN loss formulation.Source: 2) For RecCNN training · p. 5
• The residual form in Eq. (15) is described as essentially identical to Eq. (5).res(·) represents the residual learned by RecCNN.Source: 2) For RecCNN training · p. 5

IV. EXPERIMENTS

• The framework is compared against JPEG, JPEG 2000, and BPG with state-of-the-art deblocking or denoising methods.Source: §IV · p. 5
• Training uses 400 images of size 180 × 180, while testing uses 7 images excluded from the training dataset.Source: A. Datasets for Training and Testing · p. 5
• A total of 204800 patches are sampled from training images and their flip and rotation augmentations.Each image contributes 8 augmentations and 64 patches of size 40×40.Source: A. Datasets for Training and Testing · p. 5

C. Experimental Results

• At JPEG QF = 5, the framework gains 1.20dB PSNR and 0.0227 SSIM over Zhang’s.Zhang’s is identified as state-of-the-art among the compared methods.Source: C. Experimental Results · p. 6
• Removing either ComCNN or RecCNN produces worse performance than using both networks collaboratively.The ablation compares RecCNN alone with ComCNN alone plus bicubic interpolation.Source: C. Experimental Results · p. 6
• Across 0.1–0.4bpp, the framework gains average PSNR values of 3.06dB, 2.45dB, 1.34dB, and 1.09dB over JPEG2000.The corresponding average SSIM gains are 0.1047, 0.0709, 0.0525, and 0.0435.Source: C. Experimental Results · p. 6
• Against BPG, the framework gains average 0.99dB PSNR and 0.0218 SSIM while saving 5.22% bit-rates.The paper notes that improvements over BPG are less obvious because BPG is already a strong compression method.Source: C. Experimental Results · p. 7
• On Set5, Set14, LIVE1, and General-100, the proposed compression performance exceeds JPEG and JPEG2000 by a larger margin.Source: C. Experimental Results · p. 7

Runtime comparison

• The proposed compression framework runs faster than other post-processing methods.GPU runtimes could not be tested for all competing methods.Source: D. Running Time · p. 7

Overall outcome

• The framework’s conclusion reports state-of-the-art performance together with substantially faster processing than most post-processing algorithms.This summarizes the paper’s reported quality and efficiency outcome without adding a new benchmark.Source: §V · p. 7