Source-linked AI summary
Classification of Time-Series Images Using Deep Convolutional Neural Networks
Nima Hatami, Yann Gavet, Johan Debayle
TL;DR
Time-series classification commonly relies on one-dimensional signals and hand-crafted representations. The paper transforms signals into recurrence-plot texture images and classifies them with a CNN that learns features and classification jointly, achieving competitive performance against deep and state-of-the-art TSC methods.
Problem
Time-series classification has largely focused on 1D signals, limiting access to image-like texture features and unified representation learning.
Method
The paper encodes time-series as recurrence-plot gray-level texture images and applies a deep CNN that jointly learns hierarchical features and classification.
Results
The proposed model achieves 10 wins out of 20 datasets and an average rank of 2.15, outperforming the compared algorithms by the paper’s evaluation criteria.
Takeaways & Limitations
The unified RP-CNN pipeline provides a competitive approach to TSC against traditional texture-feature methods, other CNN-based methods, and state-of-the-art TSC algorithms.
Takeaways & Limitations
Deeper CNN architectures require large datasets, motivating further investigation for more difficult tasks and small sample sizes.
Abstract
from arXiv · showhide
Convolutional Neural Networks (CNN) has achieved a great success in image recognition task by automatically learning a hierarchical feature representation from raw data. While the majority of Time-Series Classification (TSC) literature is focused on 1D signals, this paper uses Recurrence Plots (RP) to transform time-series into 2D texture images and then take advantage of the deep CNN classifier. Image representation of time-series introduces different feature types that are not available for 1D signals, and therefore TSC can be treated as texture image recognition task. CNN model also allows learning different levels of representations together with a classifier, jointly and automatically. Therefore, using RP and CNN in a unified framework is expected to boost the recognition rate of TSC. Experimental results on the UCR time-series classification archive demonstrate competitive accuracy of the proposed approach, compared not only to the existing deep architectures, but also to the state-of-the art TSC algorithms.
1. INTRODUCTION
The paper frames time-series classification as recognition of 2D texture images generated from time-series, using CNNs to learn representations and classifications jointly.
- Time-series classification assigns class labels to ordered measurement sequences and applies across biomedical, financial, industrial, biometric, video, music, forecasting, and weather domains.
- Existing methods differ by feature domain, including frequency-domain and time-domain analyses, and by instance-based versus feature-based classification strategies.
- CNNs learn hierarchical features from raw inputs without hand-crafted features while jointly learning feature representation and classification.
- The proposed approach encodes time-series signals as recurrence-plot images and treats classification as texture recognition with a deep CNN.
2. RELATED WORK
Prior TSC work applies CNNs either directly to one-dimensional signals or after transforming signals into two-dimensional representations. This paper distinguishes its RP-CNN framework from hand-crafted texture features and related image encodings.
- Deep-learning applications to TSC use either modified CNNs with 1D inputs or transformations of 1D signals into 2D matrices before CNN classification.
- Related image-based approaches encode time-series with Gramian Angular Fields or Markov Transition Fields and classify the resulting images using a Tiled CNN.
- RP-based alternatives extract Gabor or Local Binary Pattern texture descriptors and use SVM classifiers in a traditional, separated feature-extraction framework.
- The proposed CNN framework automatically learns classification-useful texture features jointly with the classifier rather than relying on hand-crafted features.
3. METHODOLOGY
The methodology converts recurrent phase-space behavior into gray-level recurrence-plot images and classifies them with a two-stage CNN whose feature learning and classifier are trained together.
- 3.1 Time-Series to Image Encoding: Recurrence plots visualize recurrences of an m-dimensional phase-space trajectory through a two-dimensional representation.
- 3.1 Time-Series to Image Encoding: The recurrence matrix marks whether pairs of states are within threshold distance, while its textures and typological patterns encode structures such as periodicity, drift, and laminar states.
- 3.1 Time-Series to Image Encoding: The method skips recurrence-matrix thresholding and uses gray-level texture images to avoid information loss from binarization.
- 3.2 CNN for Time-Series Image Classification: The CNN uses two feature-learning stages combining convolution, ReLU activation, and pooling, followed by a fully connected classification layer.
- 3.2 CNN for Time-Series Image Classification: Stochastic gradient descent updates parameters through feedforward propagation and error back-propagation during training.
4. EXPERIMENTS
Experiments evaluate the proposed CNN pipeline on 20 selected UCR datasets, using error-rate comparisons with established TSC algorithms and CNN-based alternatives. The evaluation also describes the CNN configuration and kernel visualizations.
- Datasets: The evaluation uses 20 selected datasets from the UCR archive, with separate training and testing sets for comparable algorithm results.The broader archive contains 85 datasets spanning different class counts, training-set sizes, and time-series lengths.
- Evaluation: Performance is compared using Number of Wins and Average Rank, where stronger performance means more wins and a lower average rank.Missing error rates are normalized for each algorithm before calculating these measures.
- Visualization: Figure 4 visualizes kernels from the first and second hidden layers of a CNN trained on TwoPattern data.The visualization concerns learned representations rather than dataset-level performance.
- Baselines: The comparison includes 1-NN DTW, Fast-Shapelets, Bag-of-Patterns, SAX-VSM, MCNN, and GAF-MTF.The table reports error rates on the selected UCR datasets, with some missing entries.
- Results: The proposed model obtains 10 out of 20 wins with an average rank of 2.15, ahead of MCNN’s 6 wins out of 19 and average rank of 2.36.SAX-VSM follows with 6 wins out of 19 and an average rank of 3.0; GAF-MTF ranks fifth.
5. CONCLUSIONS
The paper concludes that recurrence plots combined with CNNs form a strong time-series classification pipeline. It also identifies deeper architectures and larger datasets as directions for harder tasks.
- Conclusions: The proposed pipeline transforms time-series signals into recurrence-plot texture images and classifies them with a deep CNN.The pipeline uses RP visualization and jointly learns feature representations and classification.
- Conclusions: The RP-CNN approach obtains better results than RP with traditional hand-crafted texture features and SVM, and than GAF-MTF images with CNN.The conclusion presents these comparisons as evidence for the proposed pipeline’s superiority.
- Future Work: Future work should investigate CNNs with more feature-representation layers for more difficult TSC tasks, preferably with more data samples.The paper also suggests ensemble learning and output coding for CNNs.