A Deep Image Compression Framework for Face Recognition

• Hundreds of millions of database face images create substantial storage and computing demands in large-scale recognition.Source: §1 · p. 1
• Image-quality enhancement alone may not optimize face-recognition performance because deblocking objectives differ from recognition objectives.Source: §1 · p. 2
• The proposed approach replaces conventional codecs with a convolutional autoencoder compression-reconstruction network combined with PNG and jointly trained with face recognition.Source: §1 · p. 2
• LFW compressed by the jointly optimized network achieves higher verification accuracy than LFW compressed by JPEG2000 and much higher accuracy than JPEG.Source: §1 · pp. 2–3

• CompNet produces a compact representation, PNG encodes it, RecNet reconstructs the image, and a face-recognition network extracts feature vectors.Source: §2.1 · p. 3
• The compression-reconstruction framework is directly connected to an existing face-recognition network for joint optimization.Source: §2.1 · p. 4
• Joint training aims to preserve useful original-image information while making reconstructed images close to the originals and initializing from pretrained networks.Source: §2.1 · p. 4

2.2.1 Network structure of CompNet and RecNet

• The proposed CompNet and RecNet are symmetrically designed as convolutional autoencoders for image compression and reconstruction.CompNet uses downsampling layers, while RecNet uses corresponding deconvolutional upsampling layers.Source: §2.2.1 · p. 4
• Three downsampling stages reduce the compact map’s height and width to 1/8 of the original, and three upsampling stages restore original dimensions.Source: §2.2.1 · p. 4
• Residual blocks are inserted between downsampling and upsampling layers, with dropout and PReLU replacing the default residual-block design and ReLU.Source: §2.2.1 · p. 4, Figure 6 · p. 5

2.2.2 Quantization and entropy coding

• Compact-map values are quantized from [-1, 1] to [0, 255], encoded with PNG, then mapped back to [-1, 1] before RecNet reconstruction.Quantization is omitted during training because it is non-differentiable, under an assumption of negligible error.Source: §2.2.2 · pp. 4–5

2.2.3 Training of CompNet and RecNet

• CompNet and RecNet are trained on 81,650 image blocks extracted from 1,633 high-quality images using MSE loss and SGD for 40 epochs.The training images are saved losslessly as PNG to avoid compression artifacts; batch size is 20 and the learning rate is fixed at 0.0001.Source: §2.2.3 · p. 5

2.3 Selection and pre-training of face recognition network

• Cosface is selected as the face-recognition model because its Sphere20 architecture and LMCL loss had achieved 99.28% verification accuracy on LFW before compression.Source: §2.3 · p. 5

2.3.1 Structure of Cosface Face Recognition Model

• The Sphere20 network structure is used for Cosface, with the feature-extraction parameters supplied pre-trained and the classifier parameters retrained.Source: §2.3.1 · p. 5, Figure 7 · p. 6
• The classifier is trained on 452,960 cropped CASIA-WebFace images from 10,575 individuals for six epochs using SGD with a decaying learning rate.Source: §2.3.1 · p. 5

2.4 Joint training of the combined network

• During joint training, CompNet’s compact representation is directly passed to RecNet, omitting quantization and entropy coding from the differentiable training path.Input pixels are normalized from [0, 255] to [-1, 1] before CompNet.Source: §2.4 · p. 6, Figure 8 · p. 6
• The combined network uses Cosface’s LMCL as its loss because face-recognition accuracy is the sole optimization objective during joint training.The reconstructed-image effect is ignored in this stage’s loss.Source: §2.4 · p. 6
• Pre-trained CompNet, RecNet, feature-extraction, and classifier parameters are imported into the combined network before joint training.Training uses CASIA-WebFace images of size 112 × 96, batch size 100, SGD, and 40 epochs with a learning rate reduced from 0.01 to 0.00001.Source: §2.4 · pp. 6–7

3.1 Image recovery effect of Initially trained CompNet and RecNet

• At equal bit rates, JPEG2000 produced the best subjective reconstruction, while initially trained CompNet and RecNet outputs were somewhat ambiguous and JPEG showed blocking artifacts.Source: §3.1 · p. 7, Figure 9 · p. 7
• The study prioritizes face-recognition accuracy over equal-bit-rate reconstruction quality and therefore does not provide detailed PSNR or SSIM comparisons.Source: §3.1 · p. 7, §3.2 · pp. 7–8

3.2 Comparison of face verification accuracy of LFW datasets compressed by different compression methods

• 89.48% verification accuracy at 1/7 dataset size exceeded JPEG2000's 85.93% and JPEG's 67.87% at the same compression ratio.The original dataset reached 99.28%; joint training improved the proposed network from 75.87% to 89.48%.Source: §3.2 · p. 8, Table 1 · p. 8
• Joint training raised verification accuracy from 75.87% to 89.48%, showing that recognition-oriented joint optimization improved the compressed system.Source: §3.2 · p. 8
• Despite worse visual restoration than JPEG2000, the jointly trained network achieved higher accuracy and avoided the discontinuous color patches observed in JPEG2000 reconstructions.The proposed 4.38MB images were reddish but retained important face parts; JPEG2000 images were 4.41MB.Source: Figure 10 · p. 8, §3.2 · p. 9
• Before joint training, the proposed network's blurry 2.79MB reconstructions still surpassed JPEG at 10.8MB and JPEG2000 at 2.78MB in verification accuracy.The paper attributes competing artifacts such as blocking and discontinuous color patches to false boundaries that affect essential face-feature extraction.Source: §3.2 · p. 9
• Removing compact-map quantization improved verification accuracy by 2%–3% compared with quantization, indicating quantization error leaves room for improvement.Source: §3.2 · p. 9, Table 2 · p. 9

• At the same compression ratio, the proposed jointly trained compression scheme achieved higher LFW face-verification accuracy than JPEG2000 and much higher accuracy than JPEG.Source: §4 · p. 9