Adaptive Modular Convolutional Neural Network for Image Recognition.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) are crucial for image recognition but face challenges like overfitting, large parameters, and slow convergence when scaled up.
• Existing large-scale models (e.g., VggNet, ResNet, GoogLeNet) improve accuracy but exacerbate these issues.
• Efficient and accurate image recognition models are essential for advancing computer vision tasks.

Purpose of the Study:

• To propose a modular CNN design that mitigates overfitting and reduces model parameters.
• To accelerate model convergence by fusing features from parallel submodules.
• To introduce an attention-based gate unit for dynamic optimization of model structure and reduction of floating-point operations (FLOPs).

Main Methods:

• A modular CNN architecture is designed with multiple parallel modules, each containing submodules that extract and fuse features.
• An attention mechanism with a gate unit is incorporated to dynamically select the optimal number of modules and network structure.
• The model's efficiency is evaluated by dynamically reducing FLOPs.

Main Results:

• The proposed modular CNN effectively addresses overfitting and reduces model parameters compared to traditional large-scale networks.
• Feature fusion from submodules accelerates model convergence.
• The model achieved high accuracy on benchmark datasets: Cats-vs.-Dogs (99.3%), 10-Monkey Species (99.26%), and Birds-400 (99.13%).

Conclusions:

• The novel modular CNN design offers a simple yet effective approach to image recognition.
• The attention-based optimization allows for dynamic structural adaptation and reduced computational cost (FLOPs).
• The model demonstrates superior performance and efficiency, outperforming established architectures on several datasets.