A context aware multiclass loss function for semantic segmentation with a focus on intricate areas and class...

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Image segmentation is crucial for machine vision systems, but model accuracy is often hindered by imbalanced datasets.
• Deep neural networks (DNNs) are widely used for image segmentation, yet struggle with data imbalances.
• Improving segmentation accuracy is critical for overall system performance.

Purpose of the Study:

• To introduce a novel loss function, SPix-WCE, designed to enhance DNN performance in image segmentation.
• To address data imbalances by focusing on complex image regions during model training.
• To improve segmentation accuracy through a superpixel-based weighting scheme.

Main Methods:

• Developed the SPix-WCE loss function utilizing the SLIC (Simple Linear Iterative Clustering) algorithm.
• Analyzed superpixels to identify and focus on complex image regions.
• Implemented a weighting scheme to adjust the influence of different image areas in the loss calculation.
• Conducted experiments using three DNN models and four imbalanced multiclass datasets.

Main Results:

• SPix-WCE loss function demonstrated superior performance compared to commonly used loss functions.
• Significant improvements were observed in Intersection over Union (IoU), F1-Score, and pixel accuracy metrics.
• The approach effectively handled various degrees of data imbalance across different datasets.

Conclusions:

• The proposed SPix-WCE loss function effectively boosts deep neural network performance in image segmentation tasks.
• SPix-WCE offers a robust solution for handling imbalanced image datasets.
• This method provides a valuable advancement for improving the accuracy and reliability of machine vision systems.