Filter pruning for convolutional neural networks in semantic image segmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are crucial for real-time applications but demand significant resources.
• Existing CNN compression methods often require extensive training and overlook data influence.
• Resource-efficient AI is vital for deployment on edge devices and real-time systems.

Purpose of the Study:

• To develop efficient and explainable methods for compressing Convolutional Neural Networks (CNNs).
• To investigate the impact of data on model compression and performance.
• To present novel pruning strategies applicable to various CNN architectures.

Main Methods:

• Proposed two filter pruning methods utilizing Principal Component Analysis (PCA) and a next-convolution influence-metric.
• Introduced a layer pruning approach specifically for U-Net architectures.
• Implemented a fine-tuning strategy to restore model generalization post-pruning.
• Developed an importance score distribution-based method using mean standard deviation and influence metrics.

Main Results:

• Achieved significant parameter and FLOPs reduction on U-Net (98.7% and 97.5%), DeepLabv3+ (46.5% and 51.9%), SegNet (72.4% and 83.6%), and VGG-16 (86.5% and 82.2%).
• Maintained or improved accuracy across all tested models and datasets, including a 0.35% gain for U-Net.
• Demonstrated the generic applicability of the pruning strategies across different CNNs and semantic segmentation tasks.

Conclusions:

• The developed pruning methods effectively reduce CNN complexity without compromising accuracy.
• Explainable AI (xAI) context provides insights into the relationship between data and model performance.
• These techniques offer a viable solution for deploying high-performance CNNs on resource-limited platforms.