Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Residual Feature-Reutilization Inception Network for Image Classification

Yuanpeng He, Wenjie Song, Lijian Li, Tianxiang Zhan, Wenpin Jiao

TL;DR

This paper addresses the need for efficient multi-scale feature extraction in image classification by introducing Residual Feature Reutilization Inception (ResFRI) and Split-ResFRI blocks. These blocks combine four parallel convolutional paths with information interaction passages and a residual connection to enlarge the receptive field while keeping parameter counts low; Split-ResFRI further reduces computation by partitioning the input channels into four groups with ratios $3/8$, $3/8$, $1/8$, and $1/8$. Across CIFAR-10, CIFAR-100, and Tiny Imagenet, the proposed methods achieve state-of-the-art or competitive accuracy for approximately equal model sizes and without extra data, often surpassing ResNet-101 and several inception-based baselines. The work offers a lightweight CNN backbone with strong multi-scale representation, suitable for deployment on resource-constrained devices and adaptable to tasks such as segmentation and object detection, with future directions focusing on more efficient fusion strategies and optimal channel splits.

Abstract

Capturing feature information effectively is of great importance in the field of computer vision. With the development of convolutional neural networks (CNNs), concepts like residual connection and multiple scales promote continual performance gains in diverse deep learning vision tasks. In this paper, we propose a novel CNN architecture that it consists of residual feature-reutilization inceptions (ResFRI) or split-residual feature-reutilization inceptions (Split-ResFRI). And it is composed of four convolutional combinations of different structures connected by specially designed information interaction passages, which are utilized to extract multi-scale feature information and effectively increase the receptive field of the model. Moreover, according to the network structure designed above, Split-ResFRI can adjust the segmentation ratio of the input information, thereby reducing the number of parameters and guaranteeing the model performance. Specifically, in experiments based on popular vision datasets, such as CIFAR10 ($97.94$\%), CIFAR100 ($85.91$\%) and Tiny Imagenet ($70.54$\%), we obtain state-of-the-art results compared with other modern models under the premise that the model size is approximate and no additional data is used.

Residual Feature-Reutilization Inception Network for Image Classification

TL;DR

This paper addresses the need for efficient multi-scale feature extraction in image classification by introducing Residual Feature Reutilization Inception (ResFRI) and Split-ResFRI blocks. These blocks combine four parallel convolutional paths with information interaction passages and a residual connection to enlarge the receptive field while keeping parameter counts low; Split-ResFRI further reduces computation by partitioning the input channels into four groups with ratios , , , and . Across CIFAR-10, CIFAR-100, and Tiny Imagenet, the proposed methods achieve state-of-the-art or competitive accuracy for approximately equal model sizes and without extra data, often surpassing ResNet-101 and several inception-based baselines. The work offers a lightweight CNN backbone with strong multi-scale representation, suitable for deployment on resource-constrained devices and adaptable to tasks such as segmentation and object detection, with future directions focusing on more efficient fusion strategies and optimal channel splits.

Abstract

Capturing feature information effectively is of great importance in the field of computer vision. With the development of convolutional neural networks (CNNs), concepts like residual connection and multiple scales promote continual performance gains in diverse deep learning vision tasks. In this paper, we propose a novel CNN architecture that it consists of residual feature-reutilization inceptions (ResFRI) or split-residual feature-reutilization inceptions (Split-ResFRI). And it is composed of four convolutional combinations of different structures connected by specially designed information interaction passages, which are utilized to extract multi-scale feature information and effectively increase the receptive field of the model. Moreover, according to the network structure designed above, Split-ResFRI can adjust the segmentation ratio of the input information, thereby reducing the number of parameters and guaranteeing the model performance. Specifically, in experiments based on popular vision datasets, such as CIFAR10 (\%), CIFAR100 (\%) and Tiny Imagenet (\%), we obtain state-of-the-art results compared with other modern models under the premise that the model size is approximate and no additional data is used.
Paper Structure (21 sections, 6 equations, 7 figures, 10 tables)

This paper contains 21 sections, 6 equations, 7 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Residual Feature-Reutilization Inception Network
  3. Introduction of Structure of ResFRI and Split-ResFRI
  4. Discussions on ResFRI and Split-ResFRI and multi-scale feature fusion strategy
  5. Other Important Settings of ResFRI and Split-ResFRI
  6. Limitations of existing methods
  7. EXPERIMENTS
  8. Implementation Details
  9. Metric of Image Classification
  10. Experiments on CIFAR-10
  11. Comparison with models using extra techniques and data
  12. Experiments on CIFAR-100
  13. Comparison with models using extra techniques and data
  14. Experiments on Tiny Imagenet
  15. Comparison with models using extra techniques and data
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Original Inception from GoogLeNet
  • Figure 2: Residual Feature-Reutilization Inception
  • Figure 3: Split-Residual Feature-Reutilization Inception
  • Figure 4: Details of Passages Between Convolution Groups. Note: The Split-ResFRI Also Adopts the Same Information Interaction Strategy as ResFRI
  • Figure 5: Loss and Accuracy of ResFRI-addition on Datasets
  • ...and 2 more figures