Detail Preservation Network (DPN)

• Detail Preservation Network (DPN) is a neural architecture that maintains high-resolution detail using a single high-fidelity processing stream and multi-scale DP-Blocks.
• It integrates specialized modules—such as SRM, MRM, FIFM, and GSM—to enhance fine structure while suppressing artifacts in biomedical segmentation and HDR imaging.
• Empirical results demonstrate that DPN achieves superior accuracy and efficiency, enabling fast, compact models that outperform traditional encoder-decoder methods.

A Detail Preservation Network (DPN) is a neural architecture specifically designed to maintain high-fidelity local spatial information and suppress artifacts during image processing tasks where detail retention is essential. Two distinct DPN formulations have appeared in the literature: (1) the efficient high-resolution vessel segmentation DPN for biomedical analysis (Guo, 2020), and (2) the advanced HDR imaging DPN employing a multi-stage detail enhancement and artifact suppression pipeline (Li et al., 7 Mar 2024). This entry synthesizes the architecture, methodology, and empirical performance of these DPN variants.

1. High-Resolution Single-Stream Architecture

The retinal vessel segmentation DPN dispenses with conventional encoder–decoder pipelines (e.g., U-Net), which downsample and upsample feature maps, inevitably sacrificing spatial fidelity. Instead, this DPN processes the image as a single high-resolution feature stream throughout the network (Guo, 2020). All layers, from input to output, preserve the $H\times W$ spatial dimensions of the input, ensuring preservation of sub-pixel boundary details—critical for segmenting fine structures such as one-pixel-wide retinal capillaries.

The architecture commences with a $3\times 3$ convolution (stride 1, 32 channels), followed by a cascade of 8 identical Detail-Preserving Blocks (DP-Blocks), each preserving spatial dimensions. The final output is produced via a $1\times 1$ convolution and sigmoid activation to provide a per-pixel probability map.

2. Detail-Preserving Block (DP-Block) Design

The DP-Block underpins multi-scale contextual aggregation at high spatial resolutions. For an input tensor $X\in\mathbb{R}^{H\times W\times C_{in}}$, the block generates three scale branches:

• Branch 1 (OS=1): $x_1 = \mathrm{ReLU}(W_1 * X + b_1)$, using $3\times3$ convolutions, $C_0=16$ filters.
• Branch 2 (OS=2): Downsamples $X$ via $2\times2$ max-pooling, applies $3\times3$ convolutions ($C_1=8$ filters), and then up-samples.
• Branch 3 (OS=4): Downsamples $X$ via $4\times4$ max-pooling, $3\times3$ convolutions ($C_2=8$ filters), and up-samples twice.

Features from the branches are fused in a cascaded fashion via spatial upsampling (transposed convolution) and concatenation. The final output $Y$ retains the original input’s spatial dimensions, enhancing the effective receptive field without resolution loss. Formally: $$\begin{aligned} x_1 &= \sigma(W_1 * X + b_1) \ x_2 &= \sigma(W_2 * \mathrm{Pool}_2(X) + b_2) \ x_3 &= \sigma(W_3 * \mathrm{Pool}_4(X) + b_3) \ x_4 &= \sigma(W_4 * [x_2, \mathrm{Deconv}_2(x_3)] + b_4) \ x_5 &= \sigma(W_5 * [x_1, \mathrm{Deconv}_2(x_4)] + b_5) \ Y &= \sigma(W_6 * [x_5, X] + b_6) \end{aligned}$$ with $\sigma(\cdot)$ denoting ReLU.

The effective receptive field of a neuron in $Y$ is increased up to fourfold compared to $X$; stacking 8 DP-Blocks provides an exponential receptive field growth without spatial detail degradation.

3. High Dynamic Range (HDR) Imaging Detail Preservation Pipeline

The HDR imaging DPN (Li et al., 7 Mar 2024) extends the concept of detail preservation to multi-exposure fusion, focusing on oversaturated regions and ghosting artifact suppression. The architecture incorporates:

• SHDR-ESI branch: Single-frame HDR reconstruction using a reference image and an “enhanced stop image” (ESI), with a detail enhancement mechanism (DEM) that aggregates fine structure.
• SHDR-A-MHDR branch: Multi-exposure HDR synthesis, fusing features from multiple LDR images through a Feature Interaction Fusion Module (FIFM) and purifying the result using a Ghost Suppression Module (GSM) guided by the ghost-free SHDR-ESI feature.

Detail Enhancement Mechanism (DEM)

The DEM sequentially applies:

• Self-Representation Module (SRM): Cross-attention between reference and ESI features, followed by dynamic modulation.
• Mutual-Representation Module (MRM): Modulated mutual cross-attention between reference/ESI, injecting complementary detail, followed by aggregation: $$F_s^1 = \mathrm{Conv}_{1\times1}\left(\left[E_r([F_2',F_{2,m}',\tilde F_2,\tilde F_{2,m}]), F_2\right]\right)$$ where $F_2, F_{2,m}$ are reference and ESI features, $E_r$ a small extraction head.

Repeat application of DEM stages produces deep, detail-rich representations.

Feature Interaction Fusion and Ghost Suppression

FIFM merges features from multiple exposures using cross-concatenation, multi-scale extraction, and spatially adaptive softmax attention for pixelwise fusion: $$F_{2,i} = \mathrm{Conv}_{1\times1}\left(W_{2,i}\odot F_{2,i}^{\mathrm{Cat}} + W_i\odot F_i^{\mathrm{Cat}}\right)$$ GSM deploys cross-attention—using the ghost-free SHDR-ESI feature to modulate and purify the fused multi-exposure features.

4. Training Protocols and Loss Functions

Retinal Vessel DPN

DPN is trained end-to-end from scratch using empirically balanced cross-entropy: $$L(p,y\,|\,\theta) = -\beta \sum_{y_j=1} \log p_j - (1-\beta) \sum_{y_j=0} \log(1-p_j)$$ with class-balancing weight $\beta = N_{-}/(N_{+}+N_{-})$ and auxiliary losses attached after intermediate DP-Blocks to stabilize gradients. Optimization uses Adam, batch size 1, and data augmentation via random rotations and flips.

HDR DPN

The total loss is a weighted sum of the multi-exposure and single-frame HDR branches, each incorporating reconstruction ($L_1$ in tone-mapped space), SSIM, and edge-preserving gradient differences: $L = L_M + \lambda L_S$ with hyperparameters $\lambda=0.5$, $\alpha=0.2$, $\beta=0.5$. Standard optimizer is Adam with learning rate decay.

5. Empirical Results and Ablation Studies

Retinal Vessel Segmentation

Dataset	Sensitivity	Specificity	Accuracy	AUC	F1	Speed (fps)	Model size
DRIVE	0.7934	0.9810	0.9571	0.9816	0.8289	11.8	120k
CHASE_DB1	0.7839	0.9842	0.9660	0.9860	0.8124	5.6	120k
HRF	0.7926	0.9764	0.9591	0.9697	0.7835	–	120k

DPN demonstrates segmentation accuracy comparable or superior to prior methods while being 20–160$\times$ faster and more compact (0.12M parameters vs. 1–20M for others) (Guo, 2020).

HDR Imaging

Ablation on Kalantari's HDR set shows the PSNR drops by 0.22 dB without SRM, by 0.21 dB without MRM, by 0.55 dB without FIFM, and by 0.51 dB without GSM, confirming all modules’ contributions (Li et al., 7 Mar 2024). The network yields PSNR-$\mu$ 44.39 dB, SSIM-$\mu$ 0.9915, and HDR-VDP-2 65.21, all at or above prior state-of-the-art.

6. Implementation Considerations and Best Practices

• For vessel segmentation, Xavier initialization and ReLU after every convolution are required; auxiliary losses after intermediate DP-Blocks significantly boost gradient flow. No batch normalization is used to accommodate small batch sizes.
• Data augmentation is critical due to high spatial resolution and limited data, especially for biomedical segmentation (Guo, 2020).
• For HDR synthesis, all learning and evaluation occurs in the LDR tone-mapped domain via $\mu$-law; module-specific training protocols and careful multi-stage feature fusion are necessary. PyTorch is the common framework, supporting end-to-end training (Li et al., 7 Mar 2024).

7. Context, Generalization, and Impact

DPNs empirically demonstrate that single-stream, full-resolution representations and intra-block multi-scale aggregation—eschewing downsampling bottlenecks—can preserve fine structures in domains with thin or low-contrast features. In HDR imaging, the addition of advanced feature fusion and artifact suppression (FIFM, GSM) shows state-of-the-art quantitative and perceptual metrics, especially in oversaturated or dynamic settings where prior architectures suffered from blurring or ghosting.

A plausible implication is that DPN concepts are widely transferable to other domains where structural detail and artifact suppression are bottlenecks, such as medical segmentation for very small objects, and image restoration in adverse photometric conditions.

Key References:

• "DPN: Detail-Preserving Network with High Resolution Representation for Efficient Segmentation of Retinal Vessels" (Guo, 2020)
• "Single-Image HDR Reconstruction Assisted Ghost Suppression and Detail Preservation Network for Multi-Exposure HDR Imaging" (Li et al., 7 Mar 2024)