Related Experiment Video
Updated: Jul 4, 2025
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.2K
IremulbNet: Rethinking the inverted residual architecture for image recognition
Tiantian Su1, Anan Liu1, Yongran Shi1
1Shaanxi Normal University, Xi'an 710119, Shaanxi, China.
Summary
A new lightweight Convolutional Neural Network (CNN) model, IremulbNet, improves image classification accuracy on mobile devices. This efficient network achieves higher accuracy and faster inference times compared to existing models.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Growing demand for running complex Convolutional Neural Network (CNN) models on resource-constrained mobile devices.
- Existing lightweight network models often struggle with achieving sufficient classification accuracy.
Purpose of the Study:
- To propose an efficient and lightweight neural network model, IremulbNet, to address the accuracy limitations of current mobile-optimized CNNs.
- To enhance classification accuracy while maintaining computational efficiency for mobile applications.
Main Methods:
- Introduction of an Inverse Residual Multi-Branch Network (IremulbNet) architecture.
- Reconstruction of the core module using an inverse residual structure.
- Integration of multi-branch feature extraction, a novel feature fusion method, and depthwise separable convolutions.
Main Results:
- IremulbNet demonstrated superior performance on the fine-grained Imagenet-woof dataset, outperforming MobileNet V3, ShuffleNet V2, and PeleeNet by 10.9%, 12.2%, and 15.3% in accuracy, respectively.
- Significant reduction in inference time: 42.09% faster than ResNet50 and 75.56% faster than DenseNet121 on GPU.
Conclusions:
- IremulbNet effectively balances high classification accuracy with computational efficiency for mobile deployment.
- The proposed architecture offers a promising solution for deploying advanced CNN models on mobile devices.
