PaperlayerView exact revision on arXiv

Source-linked AI summary

PP-Net: A Hybrid Physical-Prior Neural Network for Scattered Light Removal in Biomedical Images on Embedded Devices

Yongfei Guo, Tingjin Chu, Mengzhuo Liu, Hongwei Lou, Yuanhao Gong

arXiv:2609.26474v1cs.CVcs.AIcs.LGeess.IVeess.SP

TL;DR

Biomedical scattered-light removal is constrained by limited paired real clinical data and the computational demands of edge deployment. PP-Net combines noise suppression, physical-prior estimation, and neural refinement; on RK3588, inference averaged approximately 200 ms per 512×512 image.

Abstract

from arXiv · show

Scattered light is common in biomedical images, yet its removal remains challenging. The difficulty arises from three aspects: first, aligned scattered-light-free biomedical ground truth is often unavailable; second, scattering is coupled with weak illumination and sensor-induced noise; and third, many learning-based restoration models are computationally expensive for embedded devices in Internet of Medical Things (IoMT) scenarios. To address these issues, this paper proposes PP-Net, a hybrid physical-prior neural network for biomedical scattered light removal. The proposed method consists of three components: DFN-Net suppresses sensor-induced noise, ASAP estimates the scattering map and recovers a physics-based prior map, and GF-Net refines the prior map by fusing it with the denoised observation. To reduce the dependence on paired biomedical ground truth, a progressive synthetic training and cross-domain transfer strategy is developed. Experiments show that the physical-prior branch improves the peak signal-to-noise ratio (PSNR) by up to 1.26 dB on paired synthetic benchmarks. Under joint noise-and-scattering degradation, PP-Net improves PSNR by more than 10.8 dB and the structural similarity index measure (SSIM) by more than 0.62 compared with representative baseline methods. On real W2S biomedical images, the proposed method reduces the average Natural Image Quality Evaluator (NIQE) score by 43.3\%. Edge deployment with RKNN conversion and INT8 quantization achieves an average inference latency of approximately 200 ms per $512\times512$ image over 360 test images. These results demonstrate that PP-Net provides an effective and deployable solution for microscopic imaging, endoscopic inspection, and edge-assisted biomedical analysis in IoMT scenarios.

I. INTRODUCTION … D. Biomedical Restoration on Embedded Devices

Biomedical scattered-light removal is difficult because tissue scattering interacts with weak illumination and sensor noise, while aligned clean targets are scarce and IoMT deployment requires efficient, interpretable restoration. The section motivates combining physical priors and deep refinement in an edge-oriented framework.

I. INTRODUCTION

  • Biomedical imaging in IoMT uses microscopes and endoscopes for clinical observation and remote diagnosis, but scattering, weak illumination, and sensor noise can obscure diagnostically relevant structures.Practical restoration must also reduce reliance on paired biomedical ground truth and support efficient deployment on resource-constrained edge devices.Source: §I · p. 1

A. From Natural Scattering to Biomedical Imaging

  • Unlike natural-scene dehazing, biomedical scattered-light removal must account for tissue morphology, weak illumination, sensor noise, device-dependent acquisition, and generally unavailable aligned scattering-free ground truth.Natural-scene physical priors such as the dark channel prior offer interpretable scattering estimates but cannot be transferred directly without accounting for these differences.Source: A. From Natural Scattering to… · p. 1

B. Physical Priors for Scattered-Light Removal

  • Fixed-window physical priors can leak across tissue boundaries, shift edges, and be corrupted by sensor noise or heterogeneous textures, motivating structure-adaptive scattering estimation.Physical-prior methods provide interpretable intermediate estimates and can guide restoration when clean targets are unavailable.Source: B. Physical Priors for Scattered-Light… · pp. 1–2

C. Deep Restoration Under Biomedical Constraints

  • Deep restoration can recover complex textures and refine physical estimates, but biomedical use is constrained by scarce paired targets, synthetic-to-real domain gaps, and risks of structurally inconsistent details without physical constraints.These constraints motivate combining physical-prior guidance with deep restoration.Source: C. Deep Restoration Under Biomedical… · p. 2

D. Biomedical Restoration on Embedded Devices

  • Edge-side restoration can reduce data transmission and support privacy and low-delay enhancement, yet computationally intensive models limit deployment on resource-constrained devices.Lightweight models often target natural images rather than jointly addressing biomedical scattering, sensor noise, limited paired supervision, and physical interpretability.Source: D. Biomedical Restoration on Embedded… · p. 2

E. Motivations and Contributions · II. HYBRID PHYSICAL-PRIOR NETWORK FRAMEWORK · A. System-Level Architecture and Pipeline Overview

PP-Net addresses biomedical scattered-light removal by combining noise suppression, physics-based prior recovery, and learned refinement, with training and deployment strategies for IoMT constraints. Its cascaded pipeline supports flexible edge or server execution.

E. Motivations and Contributions

  • PP-Net targets the combined challenges of scattering, sensor noise, limited paired supervision, and edge constraints, since physical priors are noise-sensitive and deep models can be difficult to deploy.Deep restoration models also usually require paired clean targets.Source: E. Motivations and Contributions · p. 2
  • The proposed noise-aware degradation model decomposes biomedical image degradation into tissue scattering and sensor-induced noise.Source: E. Motivations and Contributions · p. 2
  • Progressive synthetic training and cross-domain transfer reduce dependence on paired biomedical ground-truth data.Source: E. Motivations and Contributions · p. 2

II. HYBRID PHYSICAL-PRIOR NETWORK FRAMEWORK

  • DFN-Net suppresses sensor noise, ASAP estimates scattering and recovers a physics-based prior map, and GF-Net fuses that map with the denoised observation for final refinement.This progressive design improves interpretability and reduces the learning burden on the final refinement stage.Source: §II · p. 2, E. Motivations and Contributions · p. 2, Figure 1 · p. 3

A. System-Level Architecture and Pipeline Overview

  • PP-Net combines DFN-Net, ASAP, and GF-Net, while PP-NetP denotes the ASAP-and-GF-Net physical-prior branch; clean synthetic restoration uses PP-NetP, whereas noisy synthetic and biomedical transfer use complete PP-Net.The pipeline maps a degraded input through a denoised observation, scattering map, and prior map to the final restored image.Source: §II · p. 2, A. System-Level Architecture and… · p. 4
  • The framework supports local RK3588 edge-node enhancement or remote-server collaborative medical analysis, and its implementation on a resource-constrained edge device is reported as feasible for IoMT enhancement.Source: E. Motivations and Contributions · p. 2, A. System-Level Architecture and… · p. 4

B. Biomedical Scattered Light Model · C. Initial Noise Suppression via DFN-Net · D. Advanced Scattering Adaptive Prior via ASAP

PP-Net models biomedical degradation as scattering combined with sensor noise, then uses DFN-Net to denoise before ASAP estimates scattering and recovers a physics-based prior for GF-Net refinement. ASAP’s adaptive neighborhoods are designed to improve scattering estimation while preserving local structure.

B. Biomedical Scattered Light Model

  • The task-oriented degradation model separates additive sensor noise from multiplicative scattering corruption, motivating progressive restoration.It represents scattering with S(x, y) = 1 − t(x, y).Source: B. Biomedical Scattered Light Model · p. 4
  • The model assumes localized active illumination rather than global ambient light, and introduces a scaling coefficient α for illumination and attenuation compensation.Source: B. Biomedical Scattered Light Model · p. 4
  • Sensor noise is represented analytically as generic additive noise, while noisy synthetic training uses mixed Poisson–Gaussian noise to approximate biomedical acquisition.Source: B. Biomedical Scattered Light Model · p. 4

C. Initial Noise Suppression via DFN-Net

  • Because sensor noise can distort local statistics and undermine prior estimation, DFN-Net denoises the observation before scattering estimation.Source: C. Initial Noise Suppression via DFN-Net · p. 4
  • DFN-Net uses a fully convolutional FFDNet design with dilated blocks at factors 1, 2, 4, and 8 and residual learning to focus on noise.The dilated design expands the receptive field with limited computational overhead, and the reconstructed denoised image is passed to prior estimation.Source: C. Initial Noise Suppression via DFN-Net · p. 4

D. Advanced Scattering Adaptive Prior via ASAP

  • ASAP evaluates multiple window sizes and nine full, half, or quarter spatial patterns per pixel, selecting a local neighborhood to improve scattering estimation and preserve boundaries.This adaptive mechanism addresses boundary leakage and structural artifacts associated with fixed-window dark-channel estimation.Source: D. Advanced Scattering Adaptive Prior… · pp. 4–5
  • After total-variation-guided refinement, ASAP recovers a physically interpretable prior map that removes dominant scattering, which GF-Net fuses with the denoised observation.The scaling coefficient α is adaptively determined to keep the prior within a physically meaningful intensity range and avoid numerical instability.Source: D. Advanced Scattering Adaptive Prior… · p. 5

E. Guided Fusion via GF-Net

GF-Net refines the physics-based prior map to recover subtle local details that may remain missing in complex biomedical tissues. It fuses the prior with the denoised observation using an efficient encoder–decoder design.

E. Guided Fusion via GF-Net

  • GF-Net concatenates the denoised observation and prior map, combining retained texture cues with scattering-suppressed structural guidance for refinement.This stage addresses the prior map’s limited ability to restore subtle local details in complex biomedical tissues.Source: E. Guided Fusion via GF-Net · p. 5
  • For computational efficiency, GF-Net uses NAFB blocks with SimpleGate channel interactions and simplified channel attention based on global average pooling and a lightweight 1 × 1 convolution.SimpleGate supplies nonlinearity, while SCA recalibrates channel-wise feature importance.Source: E. Guided Fusion via GF-Net · p. 5
  • GF-Net uses a four-level encoder–decoder with skip connections and a global residual connection to support multi-scale fusion and learn residual corrections.The learned nonlinear restoration function is denoted FGF(·).Source: E. Guided Fusion via GF-Net · p. 5

F. Progressive Synthetic Training and Cross-Domain Transfer Strategy · 1) Stage I: Synthetic Training of PP-NetP:

Because paired scattering-free clinical ground truth is generally unavailable, PP-Net uses progressive synthetic training before transferring the trained model to real biomedical images. Stage I trains the physical-prior branch on paired synthetic data for scattering-dominated restoration without explicit sensor noise.

F. Progressive Synthetic Training and Cross-Domain Transfer Strategy

1) Stage I: Synthetic Training of PP-NetP:

2) Stage II: Synthetic Training of PP-Net Under Joint Noise-and-Scattering Degradation: … A. Experimental Setup

PP-Net is trained progressively, adding mixed Poisson–Gaussian noise in Stage II before transferring the complete model to real biomedical images without supervised adaptation. Experiments assess its prior estimation, paired-synthetic restoration, robustness to joint degradation, and real-image transfer using reference-based or no-reference metrics as appropriate.

2) Stage II: Synthetic Training of PP-Net Under Joint Noise-and-Scattering Degradation:

  • Stage II trains the complete PP-Net on ITS images with mixed Poisson–Gaussian noise, combining denoising, scattering-map and physics-based-prior estimation, and prior refinement to jointly address noise and scattering.The mixed noise model approximates photon fluctuations and sensor noise in biomedical imaging.Source: 2) Stage II: Synthetic Training of… · p. 6

3) Cross-Domain Transfer to Real Biomedical Images:

  • After progressive two-stage training, the complete PP-Net transfers to real biomedical images for inference and evaluation without paired-ground-truth supervised adaptation.The first stage establishes prior-guided restoration with PP-NetP; the second adds DFN-Net to improve robustness under joint noise and scattering.Source: 3) Cross-Domain Transfer to Real… · p. 6

III. EXPERIMENTS

  • The experiments evaluate ASAP on multiple dehazing benchmarks, compare hybrid restoration on paired synthetic data, test joint-degradation robustness, and assess transfer to real biomedical images.Real-image assessment uses qualitative analysis and no-reference quality assessment after progressive training on synthetic hazy and noisy-hazy data.Source: §III · p. 6

A. Experimental Setup

  • ASAP is evaluated against representative physics-based dehazing methods on SOTS, HSTS, I-HAZE, O-HAZE, and D-HAZY, spanning synthetic and real haze and indoor and outdoor scenes.Source: A. Experimental Setup · p. 6
  • PP-NetP is evaluated against representative learning-based dehazing models under paired synthetic supervision on RESIDE-ITS and RESIDE-6K.Source: A. Experimental Setup · p. 6
  • The noisy robustness evaluation injects mixed Poisson–Gaussian noise into hazy images using λp = 30 and σg = 0.05 to model coexisting scattering and sensor noise.Source: A. Experimental Setup · p. 6
  • Real biomedical evaluation applies the trained model across image categories without target-domain fine-tuning or paired scattering-free ground truth; PSNR and SSIM score paired data, while lower NIQE and BRISQUE indicate better real-image quality.Source: A. Experimental Setup · p. 6

1) Implementation Details: … B. Evaluation of the Proposed ASAP

The section reports PP-Net’s experimental setup and transfer protocol, then evaluates ASAP against physics-based methods using benchmark and visual comparisons.

1) Implementation Details:

  • Experiments used four NVIDIA RTX 6000 Ada GPUs, an Intel Core i9-14900K, 256 × 256 patches, batch size 32, Adam, and 500 epochs.The initial learning rate was 1 × 10−4 with a CosineAnnealingLR schedule; the model was implemented in PyTorch.Source: 1) Implementation Details · pp. 6–7

B. Evaluation of the Proposed ASAP

2) Baseline Methods:

  • ASAP was compared with 12 representative physics-based dehazing methods.These were FVR, DCP, CAP, MR, CEP, CC, NLBF, SLP, ROP+ [27], ALSP, GLP, and IHDCP.Source: 2) Baseline Methods · p. 7
  • The hybrid framework was compared with learning-based dehazing methods, while noisy-input tests used public models without additional retraining unless specified.W2S comparisons additionally included DCP, HDCP, QDCP, and Dark; learning baselines were MSCNN, AOD-Net, GFN, MSBDN, PFDN, FFA-Net, TBN, CDVA, IDB, and MPMF-Net.Source: 2) Baseline Methods · p. 7

3) Training and Transfer Setting:

  • Training proceeded from paired synthetic hazy data without explicit sensor noise to noisy synthetic data, followed by direct cross-domain inference on real biomedical images.PP-NetP first learned physical-prior-guided restoration, then complete PP-Net was trained for joint noise-and-scattering degradation.Source: 3) Training and Transfer Setting · p. 7

B. Evaluation of the Proposed ASAP

  • ASAP better preserves structural boundaries, suppresses halo-like artifacts, and restores more natural contrast than existing prior-based methods.These visual comparisons are consistent with the quantitative results in Table I.Source: B. Evaluation of the Proposed ASAP · p. 7

C. Performance of PP-NetP on RESIDE-ITS and RESIDE-6K · D. Robustness of PP-Net Under Noisy Degradation · E. No-Reference Evaluation on Real Biomedical Images

PP-NetP achieves the best PSNR and SSIM on paired synthetic haze benchmarks, while PP-Net outperforms compared methods under joint haze-and-noise degradation. On real W2S biomedical images, it has the best NIQE scores across four averaging levels and improves structural visibility and fine detail.

C. Performance of PP-NetP on RESIDE-ITS and RESIDE-6K

  • Over 500 epochs, training and validation losses decrease rapidly early on and then converge smoothly, with validation following a similar trend.The paper interprets the curves as indicating stable optimization and generalization on both datasets.Source: C. Performance of PP-NetP on… · p. 7, Figure 5 · p. 8
  • On RESIDE-ITS and RESIDE-6K, PP-NetP achieves the best PSNR and SSIM on both datasets.The paper attributes this result to combining scattering-map estimation, prior-map recovery, and learning-based detail reconstruction under paired synthetic supervision.Source: C. Performance of PP-NetP on… · p. 7, Table II · p. 8
  • PP-NetP produces clearer structures, more faithful textures, and fewer residual haze artifacts than competing dehazing networks, especially in dense haze regions.Representative examples are shown for ITS and RESIDE-6K.Source: C. Performance of PP-NetP on… · p. 7, Figure 6 · p. 9

D. Robustness of PP-Net Under Noisy Degradation

E. No-Reference Evaluation on Real Biomedical Images

  • Across avg1, avg4, avg16, and avg400, PP-Net achieves the best NIQE scores, while BRISQUE results vary across averaging levels.Lower NIQE and BRISQUE values indicate better perceptual quality; because paired scattering-free ground truth is unavailable, the evaluation compares PP-Net with degraded inputs and prior-based baselines and considers qualitative evidence too.Source: E. No-Reference Evaluation on Real… · pp. 8–9
  • On real W2S images at avg1, avg4, avg16, and avg400, PP-Net improves structural visibility and preserves finer biomedical details, particularly in highlighted regions.The comparisons include degraded inputs and selected baselines.Source: E. No-Reference Evaluation on Real… · p. 9

F. Ablation Study · 1) Pipeline-Level Ablation: · 2) Prior-Design Ablation:

Pipeline ablations find that the complete PP-Net performs best under noisy degradation, while ASAP and GF-Net together perform best on clean data. Prior-design ablations favor the proposed adaptive prior, supporting complementary roles for physical-prior estimation and neural refinement.

F. Ablation Study

  • The ablation study covers pipeline-level, prior-design, and training-strategy choices in the proposed framework.The reported findings here address pipeline-level and prior-design ablations.Source: F. Ablation Study · p. 9

1) Pipeline-Level Ablation:

  • On clean ITS data, combining ASAP and GF-Net achieves the best restoration result, with both components contributing to performance.This ablation uses PP-NetP without DFN-Net.Source: Table V · p. 11, 1) Pipeline-Level Ablation · p. 11
  • The complete PP-Net achieves the best performance on noisy ITS data, with DFN-Net stabilizing restoration and ASAP and GF-Net providing complementary benefits.The tested setting uses the complete framework under noisy synthetic degradation.Source: Table VI · p. 11, 1) Pipeline-Level Ablation · p. 11
  • Under noisy degradation, DFN-Net stabilizes restoration, while ASAP-based prior recovery and GF-Net detail refinement provide complementary benefits.These roles are reported for the complete framework evaluated on the noisy ITS dataset.Source: 1) Pipeline-Level Ablation · p. 11

2) Prior-Design Ablation:

  • Among the compared prior variants, the proposed adaptive prior provides the most effective guidance for restoration.The comparison keeps the GF-Net refinement backbone unchanged and varies only prior estimation.Source: Table VII · p. 11, 2) Prior-Design Ablation · p. 11
  • With GF-Net fixed, the prior comparison attributes gains to both neural refinement and improved ASAP-based scattering-map estimation.The result supports complementary roles for physical-prior estimation and neural refinement.Source: 2) Prior-Design Ablation · p. 11, Table VII · p. 11

3) Training-Strategy Ablation: … B. Deployment Configuration and Edge Inference Results

Progressive two-stage training provided the best reported trade-off between noisy synthetic restoration and biomedical transferability. On RK3588, INT8 PP-Net averaged approximately 200 ms per 512×512 image, and its outputs remained visually close to PyTorch results.

3) Training-Strategy Ablation:

  • Progressive two-stage training achieved the best overall trade-off between noisy synthetic restoration and biomedical transferability, bridging synthetic haze removal, noisy degradation handling, and real biomedical inference.The training-strategy ablation reports this as the overall best trade-off.Source: Table VIII · p. 11, 3) Training-Strategy Ablation · p. 11
  • Across the experiments, ASAP was effective, PP-NetP and PP-Net showed strong restoration performance, and the major design choices were necessary.This is the overall conclusion reported after the experiments.Source: 3) Training-Strategy Ablation · p. 11

IV. PP-NET ON EMBEDDED DEVICES

  • RK3588 was used as a representative embedded platform to assess whether PP-Net could be converted, quantized, and executed efficiently for edge-side biomedical image enhancement.The deployment evaluation was framed as a practical-feasibility test rather than an assessment of RK3588 as the only target hardware.Source: §IV · p. 11

A. Model Conversion and Quantization Workflow

B. Deployment Configuration and Edge Inference Results

  • Approximately 200 ms per 512×512 image was achieved by INT8 PP-Net on RK3588, averaged over 360 test images.The deployment configuration and performance are reported for the RK3588 platform.Source: Table IX · p. 12, B. Deployment Configuration and Edge… · p. 12
  • RK3588 outputs remained visually close to PyTorch outputs across W2S averaging levels, preserving main structural details and local tissue textures without obvious deployment-induced visual artifacts.This comparison assesses deployment reliability after RKNN conversion and INT8 quantization.Source: B. Deployment Configuration and Edge… · p. 12

C. Structure Visibility Analysis · V. CONCLUSION

PP-Net combines noise suppression, scattering estimation, physical-prior recovery, and refinement, with progressive synthetic training and cross-domain transfer for real biomedical images. The reported findings include stronger structural visibility after RK3588 deployment and efficient edge inference, while future work targets domain adaptation, clinical validation, and real-time video.

C. Structure Visibility Analysis

  • The structural-visibility analysis lightly smooths each W2S image with a 5 × 5 Gaussian filter at σ = 0.95 before computing Sobel gradients and Tenengrad sharpness.A higher score indicates stronger local structural sharpness after smoothing.Source: C. Structure Visibility Analysis · p. 12
  • RK3588 PP-Net achieved higher Gaussian-smoothed Tenengrad scores than degraded inputs across W2S averaging levels, supporting improved structural visibility.The scores are a no-reference indicator of local structural sharpness; the reported comparison supports the visual observation that RK3588 outputs remain close to PyTorch outputs while preserving main structural details.Source: C. Structure Visibility Analysis · p. 12, Table X · p. 12, Figure 9 · p. 12

V. CONCLUSION

  • PP-Net integrates DFN-Net, ASAP, and GF-Net in a progressive network-prior-network pipeline for noise suppression, scattering-map estimation, prior-map recovery, and refinement.Progressive synthetic training and cross-domain transfer reduce dependence on paired biomedical ground truth for real-image inference.Source: §V · p. 12
  • PP-Net showed strong restoration performance under joint noise-and-scattering degradation, while ASAP and PP-NetP were effective on benchmark datasets.The conclusion also reports promising visual enhancement and competitive no-reference image quality after direct transfer to real biomedical images.Source: §V · p. 13
  • RK3588 deployment with RKNN conversion and INT8 quantization averaged approximately 200 ms per image across 360 test images.The conclusion presents this as efficient edge-side biomedical enhancement.Source: §V · p. 13
  • Future work targets unsupervised biomedical domain adaptation, task-oriented clinical validation, and hardware-aware acceleration for real-time biomedical video enhancement.Source: §V · p. 13
Loading 2609.26474v1…