State of the Art in

Vector Image Classification

1. Introduction

As vector images are increasingly used in fields such as industrial design and technical drawings, developing effective classification and retrieval methods is crucial. I have provided some views for vector image classification by studying the existing literature, available technical libraries, and relevant datasets.

2. Literature Review

2.1 Current Research in Vector Image Classification

Currently, research on vector image classification focuses on two main approaches: feature extraction-based methods and deep learning-based methods. Below are some key findings:

Feature Extraction-Based Vector Image Classification: Early research mainly utilized traditional feature extraction techniques, such as shape descriptors, geometric features, and path analysis methods. These methods classify images by analyzing the geometric shapes and structural information in vector images. Typical features include Bezier curve parameters, path directions, lengths, and other geometric data.

Deep Learning-Based Vector Image Classification: In recent years, researchers have begun exploring the conversion of vector images to raster images (e.g., PNG or JPEG formats) and applying convolutional neural networks (CNNs) for classification. While this method has achieved significant success in image recognition, the process of converting vector images to raster may lead to the loss of important vector information.

2.2 Conversion between Vector and Raster Images

Some studies recommend converting vector images into raster images before classification to leverage existing deep learning tools and models. However, the effectiveness of this approach depends on the type of image and the application domain. In certain cases, retaining the original vector information may be more advantageous for classification tasks.

3. Available Technical Libraries

3.1 Libraries for Vector Image Processing

SVG.js and D3.js: These JavaScript libraries provide rich APIs for parsing and manipulating SVG files. They are widely used in data visualization and graphical applications and can serve as preprocessing tools for machine learning models.

Python SVGPath Library: This library can parse SVG path data and extract shape information, supporting feature-based methods.

TensorFlow Graphics and PyTorch3D: These libraries provide tools for handling and manipulating geometric shapes, although they are mainly used for 3D data. Some of their functions also apply to 2D vector images.

3.2 Tools for Vector to Raster Conversion

Cairo and ImageMagick: These tools support converting vector images into raster formats to apply traditional image classification models.

4. Relevant Datasets

Currently, there are relatively few publicly available datasets for vector image classification. Some commonly used datasets include:

SVG-Icons8 Dataset: Contains SVG format images across multiple categories, often used for object detection and classification tasks.

VectorNet Dataset: Primarily used in the autonomous driving field, containing vector format images of roads and traffic signs.

Kimia Shape Datasets: The Kimia Shapes99 and Kimia Shapes216 datasets are commonly used for shape retrieval and classification tasks. These datasets focus on object silhouettes and have been tested for robustness in classification tasks​

MPEG-7 CE Shape-1 Dataset: This dataset consists of 1,400 images across 70 object categories and is used for evaluating shape retrieval and classification methods. It is particularly useful for benchmarking feature extraction methods

5. Challenges and Gaps in Research

Limited Datasets: There are few datasets available for vector image classification, limiting the generalization capabilities of models and the scope of research.

Complexity of Feature Extraction: Vector images contain unique structural information that differs from raster images, requiring the development of specific feature extraction methods.

Challenges in Model Design: Existing deep learning models are typically optimized for raster images. Designing models suitable for vector images remains an open question.

6. Conclusion

Although current methods primarily rely on rasterization, researchers are working towards developing more direct classification techniques for vector images. Future research should focus on expanding datasets, exploring the unique features of vector images, and developing new machine learning models better suited to the characteristics of vector images.

7. References

Chen F. Bezier-based Regression Feature Descriptor for Deformable Linear Objects[J]. arXiv preprint arXiv:2312.16502, 2023.

Raj E F I, Balaji M. Shape Feature Extraction Techniques for Computer Vision Applications[M]//Smart Computer Vision. Cham: Springer International Publishing, 2023: 81-102.

Agarap A F. An architecture combining convolutional neural network (CNN) and support vector machine (SVM) for image classification[J]. arXiv preprint arXiv:1712.03541, 2017.

Lorente Ò, Riera I, Rana A. Image classification with classic and deep learning techniques[J]. arXiv preprint arXiv:2105.04895, 2021.

Chang L, Arias-Estrada M, Hernández-Palancar J, et al. An Efficient Shape Feature Extraction, Description and Matching Method Using GPU[C]//Pattern Recognition Applications and Methods: Third International Conference, ICPRAM 2014, Angers, France, March 6-8, 2014, Revised Selected Papers 3. Springer International Publishing, 2015: 206-221.