Integrating Multiple Receptive Fields Through Grouped Active Convolution.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional networks (CNNs) are pivotal in computer vision, with success attributed to architectural innovations.
• Existing convolution units possess fixed shapes, limiting their receptive fields and adaptability.
• Previous work introduced the active convolution unit (ACU) for dynamic shape definition and self-learning.

Purpose of the Study:

• To provide a detailed analysis of the active convolution unit (ACU).
• To demonstrate ACU's efficacy as a sparse weight convolution representation.
• To extend ACU to grouped and depthwise variants for enhanced receptive field processing.

Main Methods:

• Detailed analysis of the active convolution unit (ACU) properties.
• Extension of ACU to grouped (GACU) and depthwise (DACU) configurations.
• Experimental evaluation comparing ACU variants against traditional convolution methods.

Main Results:

• The active convolution unit (ACU) is confirmed as an efficient sparse weight convolution.
• Grouped ACU (GACU) maintains accuracy with fewer parameters, unlike naive grouped convolutions.
• Depthwise ACU (DACU) demonstrates efficiency and effectiveness in replacing standard convolutions.

Conclusions:

• The active convolution unit (ACU) and its variants offer significant improvements over traditional convolutional approaches.
• ACU-based methods provide a flexible and efficient alternative for various vision tasks.
• The proposed depthwise ACU (DACU) shows strong potential for practical implementation in deep learning models.