Inner-Imaging Networks: Put Lenses Into Convolutional Structure.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep convolutional networks (CNNs) achieve high performance but face challenges with computational costs and redundancies.
• Existing methods focus on filter diversity but overlook internal convolutional structure's complementarity and completeness.

Purpose of the Study:

• To propose a novel inner-imaging (InI) architecture to address the limitations of current CNNs.
• To enhance the relationships between channels, improving diversity, complementarity, and completeness within the convolutional structure.

Main Methods:

• Organizing channel signal points into groups using convolutional kernels to model intragroup and intergroup relationships.
• Mapping channel signals onto a pseudoimage, akin to adding a lens to the internal convolution structure.
• Implementing a lightweight and self-organizing strategy for CNNs.

Main Results:

• The InI architecture successfully increases channel diversity.
• Explicit enhancement of channel complementarity and completeness was observed.
• The InI mechanism demonstrated effectiveness when integrated with popular CNN backbones across benchmark datasets (CIFAR, SVHN, ImageNet).

Conclusions:

• The proposed InI architecture offers an efficient self-organization strategy for CNNs.
• InI effectively improves the performance of deep convolutional networks by addressing internal structural limitations.
• The method is lightweight, easy to implement, and validates its efficacy through extensive experiments.