Scene Segmentation with DAG-Recurrent Neural Networks.

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Scene segmentation is crucial for understanding image content.
• Existing methods like Fully Convolutional Networks (FCNs) struggle with rich contextual dependencies.
• Class imbalance in datasets poses a significant challenge for accurate segmentation.

Purpose of the Study:

• To propose a novel deep learning architecture for enhanced scene segmentation.
• To improve context aggregation capabilities over traditional CNNs.
• To address the issue of class imbalance in scene segmentation datasets.

Main Methods:

• Developed Directed Acyclic Graph-Recurrent Neural Networks (DAG-RNN) for context aggregation.
• Integrated DAG-RNN with pre-trained Convolutional Neural Networks (CNNs) to embed context into local features.
• Introduced a novel class-weighted loss function to handle imbalanced class frequencies.

Main Results:

• DAG-RNN demonstrated superior context aggregation compared to plain CNNs (FCNs).
• The proposed method achieved significant performance improvements on scene segmentation tasks.
• DAG-RNN requires fewer parameters and less computation than FCNs, making it suitable for embedded devices.

Conclusions:

• DAG-RNN offers a more effective and efficient approach to scene segmentation.
• The class-weighted loss function successfully boosts the performance on infrequent classes.
• The model shows strong potential for real-world applications, including those on resource-constrained devices.