Automatic Recognition for IoT Supervision Images Based on Modal Decomposition

Yan Wang, Yifeng Wang, Shengyu Zhang, Shimei Lin, Chao ChenTraitement 2022 · DOI

Abstract

The automatic recognition for Internet of Things (IoT) supervision images is a prerequisite for the detection of abnormalities in monitoring images. This technology is a developmental trend in video surveillance. Various video image detection and recognition methods have certain weaknesses, such as poor generalization ability and poor anti-interference ability. In response, this paper conducts a study on automatic recognition for IoT supervision images based on modal decomposition. The paper presents an overall framework of the IoT supervision system. For the problems of poor real-time performance and few samples that commonly exist in video stream target recognition, the paper proposes a dynamic modal decomposition-based feature extraction algorithm for IoT supervision video stream to build a suitable platform for IoT supervision image foreground segmentation. The paper selects a dictionary with rich elements and exchanges a high computation time for minimizing the reconstruction error generated by applying the dynamic modal decomposition method. Experimental results validate the effectiveness of the proposed algorithm.

Introduction

Traditional video surveillance is rapidly evolving into intelligent, IoT-integrated systems. However, these systems face significant hurdles: IoT cameras often produce low-quality images due to aggressive compression, high noise levels, and varying lighting conditions. Existing deep learning models frequently struggle in these environments because they require massive amounts of labeled data—the "few-sample" problem—and often exhibit poor generalization when deployed in new locations. Furthermore, many algorithms fail to extract moving targets effectively without a clean background frame. This research presents a robust approach to these challenges by leveraging Dynamic Modal Decomposition (DMD) and dictionary learning to separate moving targets from complex backgrounds in real-time.

IoT Supervision System Framework

The proposed system operates within a multi-layered architecture designed for efficient data flow from the edge to the cloud. The framework consists of three primary segments: the IoT supervision terminal, the data transmission system, and the remote monitoring cloud platform.

Overall framework of the IoT supervision system

As shown in the framework, the supervision terminal captures images and establishes a TCP connection via 4G communication. Data is transmitted through wireless base stations and the Internet to a cloud platform, where it is processed for data exchange, storage, and target recognition. This structure allows for remote monitoring via PCs and mobile devices, ensuring that supervision is accessible and scalable.

Methodology: Dynamic Modal Decomposition and Dictionary Learning

The core of the proposed solution is a multi-stage feature extraction algorithm centered on Dynamic Modal Decomposition (DMD). Unlike frame-by-frame pixel analysis, DMD treats the video stream as a dynamical system, breaking it into spatiotemporal "modes" based on frequency.

  1. Spatiotemporal Decomposition: The system separates the video into modes where low-frequency modes (eigenvalues near the origin) represent the static background, and higher-frequency modes represent moving foreground targets.
  2. Dictionary Learning: To address the few-sample problem, the system constructs a dictionary of essential features. By selecting a dictionary with rich elements, the researchers trade initial computation time for a significant reduction in reconstruction errors.
  3. Optimization: The algorithm utilizes $L_1$-parametric regularization and a fast iterative shrinkage-thresholding approach to solve sparse coding equations, allowing the system to filter noise and focus on essential dynamic features.
Flowchart of the proposed feature extraction and recognition algorithm

The flowchart above details the algorithmic process, from inputting IoT supervision video images to the final reconstruction of images with separated background and foreground structures. A unique correction matrix is also introduced to minimize the error between the reconstructed DMD model and the original image sequence.

Experimental Results and Discussion

The algorithm was evaluated using nine distinct sample sets from various surveillance locations and a specific industrial IoT pipeline dataset involving product recognition on a conveyor belt.

  • Processing Speed: The optimized recognition algorithm consistently performed in under 40ms, making it suitable for real-time edge processing. This represents a 20% speed improvement over standard self-recognition methods.
  • Recognition Accuracy: In industrial pipeline tests, the system achieved high recognition rates across different product types, ranging from 88% to 98%.
  • Robustness: The system maintained high F1 values (up to 0.795) even in scenarios with minimal background changes, proving its effectiveness in typical static surveillance environments.

One notable finding was the "static target paradox": if a target remains stationary for too long, it begins to blend into the background modes, which can reduce the F1 score. This highlights a fundamental challenge in modal-based vision regarding the distinction between stopped objects and permanent fixtures.

Conclusion

This study demonstrates that dynamic modal decomposition, combined with dictionary learning, provides a powerful framework for real-time IoT image recognition. By treating video as a series of spatiotemporal modes, the system achieves high accuracy and noise resilience even with limited training samples. Future work will focus on automating regularization parameter tuning and integrating memory components to better retain the identity of targets that become temporarily static.

Summary updated

Abstract

The automatic recognition for Internet of Things (IoT) supervision images is a prerequisite for the detection of abnormalities in monitoring images. This technology is a developmental trend in video surveillance. Various video image detection and recognition methods have certain weaknesses, such as poor generalization ability and poor anti-interference ability. In response, this paper conducts a study on automatic recognition for IoT supervision images based on modal decomposition. The paper presents an overall framework of the IoT supervision system. For the problems of poor real-time performance and few samples that commonly exist in video stream target recognition, the paper proposes a dynamic modal decomposition-based feature extraction algorithm for IoT supervision video stream to build a suitable platform for IoT supervision image foreground segmentation. The paper selects a dictionary with rich elements and exchanges a high computation time for minimizing the reconstruction error generated by applying the dynamic modal decomposition method. Experimental results validate the effectiveness of the proposed algorithm.
BibTeX
@inproceedings{wang2022automatic,
  author = {Yan Wang and Yifeng Wang and Shengyu Zhang and Shimei Lin and Chao Chen},
  title = {{Automatic Recognition for IoT Supervision Images Based on Modal Decomposition}},
  booktitle = {Traitement},
  year = {2022},
  doi = {10.18280/ts.390431},
}
Figure