Performance analysis of U-Net with hybrid loss for foreground detection

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) show promise in foreground detection.
• Existing background subtraction algorithms struggle with complex backgrounds and true foreground extraction.

Purpose of the Study:

• To propose a novel hybrid loss-assisted U-Net framework for enhanced foreground detection.
• To address the limitations of current methods in challenging environments.

Main Methods:

• Developed a deep learning model integrating transfer learning and a hybrid loss function.
• Incorporated reference background images and change detection masks into the U-Net architecture.
• Investigated the efficacy of combining two significant loss functions to mitigate class imbalance.

Main Results:

• The proposed framework demonstrates superior performance on standard datasets.
• Achieved better results than top-ranking methods in complex background scenarios.
• Validated effectiveness on unseen video data, confirming robustness.

Conclusions:

• The hybrid loss-assisted U-Net framework offers a significant advancement in foreground detection.
• The method effectively tackles class imbalance and complex background challenges.
• The approach shows strong generalization capabilities on diverse datasets and real-world scenarios.