SR Logo/AI Engineering
Back to Projects Catalog
Computer VisionCOMPLETED

Maritime Low-Visibility Video Restoration Pipeline

Video enhancement and image restoration pipeline engineered to mitigate fog, heavy haze, and low-visibility conditions in naval surveillance operations.

CategoryComputer Vision
Statuscompleted System
Core Tooling
PyTorchOpenCVCNNsPythonImage EnhancementCUDA

01. Problem & Operational Context

Naval defense and maritime surveillance cameras operate under severe visual degradation—dense marine fog, sea spray, night glare, and low illumination—which drastically impedes target detection.

02. Technical Approach

Engineered an image and video enhancement pipeline combining deep learning restoration networks with contrast-limited adaptive histogram equalization (CLAHE) and dark channel prior estimation to restore edge clarity.

03. System Architecture & Pipeline

End-to-End System Pipeline & Computation Flow

Multi-stage video restoration pipeline enhancing degraded video frames before passing to downstream object detection.

STEP 01OpenCV / NumPy
Frame Decomposition

Separating luminance and chrominance channels with noise variance estimation.

STEP 02Dark Channel Prior / Python
Atmospheric Light & Fog Estimation

Estimating transmission maps and haze thickness across the maritime horizon.

STEP 03PyTorch / CUDA
Deep Neural Edge Restoration

CNN-based feature enhancement sharpening high-frequency structural contours of distant vessels.

STEP 04CLAHE / OpenCV
Dynamic Range Recombination

Adaptive histogram equalization and color balance reconstruction for operator visibility.

04. Key Implementation Details

  • Engineered maritime video enhancement pipelines tailored specifically for low-visibility naval defense operations.
  • Implemented transmission map estimation to remove atmospheric haze and scattering artifacts.
  • Optimized enhancement kernels for low-latency batch processing on maritime surveillance hardware.
  • Delivered scalable perception preprocessing modules integrated into naval hardware streams.

05. Architectural Tradeoffs & Decisions

Decision: Combined analytical dehazing with lightweight CNN enhancement rather than heavy end-to-end generative models.
Justification

Prevents non-deterministic hallucination of marine obstacles while maintaining strict throughput for tactical operations.

Alternative Considered

Generative diffusion dehazers which hallucinated artificial texture patterns.

06. Engineering Challenges & Failures

Challenge: Over-amplification of sensor noise and chromatic noise in dark oceanic video frames.
Solution: Introduced bilateral filtering before contrast stretching to preserve vessel silhouette boundaries while smoothing sensor grain.
Challenge: Latency budget constraints on live naval camera feeds.
Solution: Processed luminance channel through neural enhancement while applying fast analytical transforms to chromatic channels.

07. Evaluation & Observations

Evaluation Methodology:

Evaluated edge sharpness, contrast improvement, and downstream detection accuracy on foggy naval test footage.

Key Empirical Finding:

Pre-processing foggy streams through the restoration pipeline significantly boosted downstream YOLO detection confidence.

Verified Artifacts
architectureMaritime Video Restoration Flow

Transmission map estimation, neural enhancement, and CLAHE dynamic range reconstruction.

Project LinksView RepositoryAll Projects