Column Row Convolutional Neural Network: Reducing Parameters for Efficient Image Processing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models achieve progress by increasing parameters, but this demands significant computing resources.
• Model compression techniques are crucial for reducing parameters while preserving performance.
• Convolutional Neural Networks (CNN) offer better efficiency than Fully Connected (FC) networks.

Purpose of the Study:

• To propose a novel Convolutional Neural Network (CNN) architecture, the Column Row Convolutional Neural Network (CRCNN).
• To significantly reduce the number of learning parameters and computational steps in deep learning models.
• To demonstrate the effectiveness of CRCNN in maintaining accuracy and its applicability in anomaly detection.

Main Methods:

• The CRCNN applies 1D convolution to image data using local receptive fields in both column and row directions.
• Feature abstraction is performed by processing data along each direction independently.
• Features from both directions are concatenated before being fed into a Fully Connected (FC) layer.

Main Results:

• The CRCNN significantly reduces the number of learning parameters and operational steps compared to traditional CNNs.
• Experimental results show comparable accuracy to existing methods with fewer parameters.
• The CRCNN architecture demonstrates feasibility for one-class anomaly detection tasks.

Conclusions:

• The proposed CRCNN offers an efficient alternative for deep learning by reducing model complexity.
• CRCNN effectively balances performance and computational cost, making it suitable for resource-constrained environments.
• The CRCNN's successful application in anomaly detection highlights its versatility for various real-world problems.