Semantic segmentation method of underwater images based on encoder-decoder architecture.

Area of Science:

• Marine Biology
• Computer Vision
• Deep Learning

Background:

• Underwater image quality is a significant challenge for deep learning-based semantic segmentation.
• Low-quality images lead to blurred edges, inaccurate segmentation, and poor boundary details.

Purpose of the Study:

• To propose an advanced semantic segmentation method for improving underwater image processing.
• To address the limitations of traditional methods in segmenting low-quality underwater imagery.

Main Methods:

• Image enhancement using multi-spatial transformation to improve original image quality.
• Densely connected hybrid atrous convolution to expand receptive fields and manage resolution reduction.
• Cascaded atrous convolutional spatial pyramid pooling for integrating multi-scale boundary features.
• Context information aggregation decoder for fusing shallow and deep network features to reduce information loss.

Main Results:

• The proposed method demonstrated superior segmentation integrity, location accuracy, and boundary clarity compared to state-of-the-art algorithms.
• Achieved the highest Mean Intersection over Union (MIOU) of 68.3 and Overall Accuracy (OA) of 79.4 on benchmark datasets.
• Exhibited low resource consumption and validated effectiveness through ablation experiments.

Conclusions:

• The developed method significantly enhances semantic segmentation of underwater images.
• It offers a robust solution for marine resource exploration and development by improving visual data analysis.
• The approach effectively overcomes the challenges posed by low-quality underwater imagery.