Dynamic Mosaic algorithm for data augmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) excel in computer vision but deeper architectures increase overfitting risk, reducing recognition accuracy.
• Existing data augmentation methods like the mosaic algorithm can lead to information loss due to gray backgrounds.
• Network overfitting is a significant challenge in deep learning models, impacting generalization performance.

Purpose of the Study:

• To improve the recognition accuracy of CNN models for image recognition.
• To overcome the overfitting problem in deep neural networks.
• To address information waste in mosaic-based data augmentation.

Main Methods:

• Proposed the Dynamic Mosaic algorithm, an enhancement of the mosaic algorithm, featuring dynamic adjustments to minimize gray background and increase spliced image count.
• Introduced a Multi-Type Data Augmentation (MTDA) strategy, leveraging the Dynamic Mosaic algorithm.
• Implemented MTDA by dividing training samples into four parts, each undergoing distinct augmentation operations to boost information variance.

Main Results:

• The Dynamic Mosaic algorithm effectively reduces information waste from gray backgrounds in augmented images.
• The MTDA strategy successfully prevents network overfitting by increasing information variance among training samples.
• Experimental results on the Pascal VOC dataset demonstrate superior recognition accuracy compared to state-of-the-art algorithms.

Conclusions:

• The Dynamic Mosaic algorithm and MTDA strategy are effective in enhancing CNN performance for image recognition.
• These novel approaches significantly improve model accuracy and mitigate overfitting issues.
• The proposed methods offer a promising solution for advancing deep learning in computer vision applications.