Diversifying Inference Path Selection: Moving-Mobile-Network for Landmark Recognition.

Area of Science:

• Computer Vision
• Deep Learning
• Mobile Computing

Background:

• Deep convolutional neural networks (CNNs) excel in computer vision but suffer from high computational costs, limiting mobile applications.
• Efficient network learning methods are crucial for deploying advanced AI on portable devices.

Purpose of the Study:

• To propose a novel efficient network architecture, M²Net (Moving-Mobile-Network), for landmark recognition.
• To enhance recognition accuracy by promoting diverse inference path selection using geographic information.

Main Methods:

• Developed M²Net, a network architecture integrating landmark images with geo-location data.
• Designed a novel reward function utilizing geo-location and landmark features to guide inference path selection.
• Constructed two landmark image datasets with associated geographic information for experimentation.

Main Results:

• M²Net demonstrated improved landmark recognition accuracy compared to existing methods.
• The proposed architecture achieved comparable computational complexity.
• M²Net architecture showed potential to improve the performance of other portable networks.

Conclusions:

• M²Net offers an effective solution for efficient and accurate landmark recognition on mobile devices.
• Integrating geo-location data with a dynamic inference path selection mechanism enhances deep learning model performance.
• The proposed approach advances the applicability of deep learning in resource-constrained environments.