ZeroDCE-U and G-UNet for Underwater Image Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Underwater surveillance faces challenges due to degraded image quality caused by underwater dynamics, low contrast, blurriness, and limited lighting, making it difficult to detect moving objects effectively.
Innovation Solution
An end-to-end image enhancement and object detection system utilizing the ZeroDCE-U architecture for image enhancement and G-UNet for moving object detection, which employs a convolutional neural network to estimate power curves and integrates a modified underwater image quality measure as a cost function to enhance image quality and detect moving objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If conventional gamma coefficient adjustment is used for image enhancement, then image contrast can be improved, but the parameter tuning process becomes cumbersome and time-consuming
Solution Approach 1:
The system employs a Convolutional Neural Network (DCE-net) that automatically learns and generates optimal power law curves for image enhancement without requiring manual gamma parameter tuning. The network processes input images and directly outputs enhancement curves, enabling the system to self-adjust to different underwater imaging conditions without human intervention.
Solution Approach 2:
The patent replaces the manual mechanical adjustment of gamma coefficients with an automated deep learning system. The Convolutional Neural Network substitutes the traditional manual parameter tuning mechanism, using learned features and patterns to automatically determine optimal enhancement parameters for each image.
2Ease of operation
If deep learning models are used to automatically generate power curves for image enhancement, then ease of operation is improved, but computational complexity and processing time increase
Solution Approach 1:
The system performs preliminary training of the Convolutional Neural Network offline using paired underwater images, where the network learns to map degraded images to their enhanced versions. Once trained, the model can quickly generate enhancement curves for new images without requiring complex real-time computations, thus reducing online computational burden.
Solution Approach 2:
The patent applies partial action by focusing the deep learning model specifically on estimating power law curves for enhancement rather than performing full image processing. This targeted approach reduces computational complexity compared to comprehensive image processing while still achieving effective enhancement through the learned curve parameters.
3Manufacturing precision
If multiple modules (UICM, UiConM, UISM) are integrated as cost functions for image enhancement, then image quality is improved, but system complexity increases
Solution Approach 1:
The patent merges three separate evaluation modules - Underwater Image Colour Measure (UICM), Underwater Image Contrast Measure (UiConM), and Underwater Image Sharpness Measure (UISM) - into a unified cost function framework. These modules work together as combined loss functions during network training, allowing the system to simultaneously optimize multiple quality aspects through a coordinated multi-objective optimization approach.
4Measurement precision
If G-UNet architecture is used for moving object detection, then detection accuracy is improved, but processing speed may be reduced
Solution Approach 1:
The G-UNet architecture segments the detection task into distinct encoder and decoder components with skip connections. The encoder extracts hierarchical features at multiple scales, while the decoder reconstructs precise object boundaries. This segmentation allows the model to process different aspects of detection (feature extraction, boundary localization) in specialized sub-networks, improving accuracy without excessive computational overhead.
Data Source
AI summary
An image enhancement and object detection system is provided to enhance the degraded images and bring out the inherent details of the image received via an underwater surveillance system. The said system includes a novel deep learning architecture i.e., a Zero-DCE-U that enhances the degraded underwater images using modified UIQM. Further, a new training method using a single image based on correlation is employed. Furthermore, a G-UNet architecture is configured to detect moving objects while preserving the spatial-contextual relationships.