Shallow-UWnet: Compressed Model for Underwater Image Enhancement (Student Abstract)

Ankita Naik, Apurva Swarnakar, Kartik MittalAAAI 2021 · DOI

Introduction

Underwater imaging is critical for marine exploration and autonomous underwater vehicles (AUVs). However, the optical properties of water—specifically light absorption and scattering—introduce significant degradation, including low contrast, color casts (typically blue or green), and blurring. While deep learning approaches like Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) have advanced the state-of-the-art in image enhancement, their high computational and memory requirements often preclude real-time deployment on resource-constrained AUV hardware. This paper introduces Shallow-UWnet, a lightweight architecture designed to provide high-quality enhancement with minimal computational overhead.

Proposed Architecture: Shallow-UWnet

Shallow-UWnet prioritizes efficiency without sacrificing performance by utilizing a shallow yet strategically connected network. The architecture departs from the "deeper is better" trend, focusing on three primary pillars:

  1. Shallow ConvBlocks: The network employs three sequential "ConvBlocks." Each block contains two stages of Convolution (3x3), Dropout, and ReLU activation. The inclusion of dropout layers is vital for regularizing the model and preventing overfitting on limited underwater datasets.
  2. Global Skip Connections: To preserve structural integrity and prevent the loss of fine details common in deep architectures, the original input image is concatenated with the output of every ConvBlock. This ensures that the network has constant access to the raw spatial information throughout the enhancement process.
  3. Hybrid Loss Function: The model is trained using a combination of Mean Squared Error (MSE) for pixel-level fidelity and VGG Perceptual Loss. The latter leverages a pre-trained VGG-19 network to ensure that enhanced images maintain high-level semantic and textural features that align with human visual perception.
Architecture of Shallow-UWnet featuring ConvBlocks and skip connections.

Experimental Setup

The performance of Shallow-UWnet was evaluated against three benchmark datasets to ensure generalization across diverse underwater conditions:

  • UIEB (Underwater Image Enhancement Benchmark): A real-world dataset with reference images created via human voting.
  • EUVP-Dark: A dataset simulating challenging, low-light underwater conditions using GAN-generated distortions.
  • UFO-120: A dataset featuring distorted images created via style transfer techniques.

The model was compared against several state-of-the-art (SOTA) baselines, including WaterNet, FUnIE-GAN, and Deep SESR.

Results and Discussion

Shallow-UWnet demonstrates a superior balance between enhancement quality and computational efficiency.

Efficiency and Speed

The model achieves significant compression compared to existing SOTA methods. Shallow-UWnet is approximately 18.17 times smaller than FUnIE-GAN and exhibits inference speeds up to 24 times faster than WaterNet. This efficiency makes it uniquely suited for real-time applications on portable, battery-powered underwater devices.

Enhancement Quality

Quantitative results across the three datasets are summarized in the table below. Shallow-UWnet consistently performs at or above the level of much larger models. Notably, on the EUVP-Dark dataset, it achieved the highest Peak Signal-to-Noise Ratio (PSNR) of 27.39 and the best Structural Similarity Index (SSIM) of 0.83. While larger models like WaterNet maintain a slight edge in certain metrics on the UIEB dataset (e.g., a UIQM of 3.02 vs. 2.77), Shallow-UWnet remains highly competitive while requiring a fraction of the resources.

Quantitative comparison of Shallow-UWnet against state-of-the-art models across three datasets.

Conclusion

Shallow-UWnet addresses the "computational tax" of modern underwater image enhancement. By combining a shallow architecture with global skip connections and perceptual loss, the model achieves state-of-the-art performance while being significantly smaller and faster than its predecessors. Future work will focus on deploying the model on embedded hardware like NVIDIA Jetson modules to measure real-world frames-per-second (FPS) and exploring further compression techniques such as quantization.

Summary updated

BibTeX
@inproceedings{naik2021shallow,
  author = {Ankita Naik and Apurva Swarnakar and Kartik Mittal},
  title = {{Shallow-UWnet: Compressed Model for Underwater Image Enhancement (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i18.17923},
}
Figure