Hybrid-Input Convolutional Neural Network-Based Underwater Image Quality Assessment

Area of Science:

• Computer Vision
• Robotics
• Image Processing

Background:

• Underwater operations require precise environmental sensing, driving interest in underwater image processing.
• Remotely operated vehicles (ROVs) generate redundant underwater images, pressuring equipment and operators.
• Selective image transmission based on degradation is needed to alleviate this pressure.

Purpose of the Study:

• To propose an end-to-end hybrid-input convolutional neural network (HI-CNN) for predicting underwater image degradation.
• To develop a model that enables selective image transmission, reducing data load.
• To establish a real-world dataset for practical underwater environments.

Main Methods:

• A feature extraction module concurrently processes original underwater images and saliency maps using two identical, shared-parameter branches.
• An end-to-end model integrates the feature extraction module with a prediction module to estimate image quality scores.
• A real-world dataset was created to validate the model in practical underwater scenarios.

Main Results:

• The proposed HI-CNN model effectively predicts the degradation of underwater images.
• Experimental results demonstrate the model's superior performance compared to existing methods.
• The model facilitates selective image transmission based on predicted quality.

Conclusions:

• The developed HI-CNN offers a robust solution for assessing underwater image quality.
• This approach can significantly enhance the efficiency and reliability of underwater operations.
• The established dataset supports further research and practical application of underwater image processing techniques.