Computer-Vision-Oriented Adaptive Sampling in Compressive Sensing.

Area of Science:

• Signal Processing
• Computer Vision
• Sensor Technology

Background:

• Compressive sensing (CS) is vital for signal compression in Internet of Things (IoT) systems, reducing transmission costs.
• Decreased sampling rates in CS degrade signal quality, impacting computer vision (CV) inference accuracy.
• Existing CS methods struggle to balance compression with the specific needs of CV tasks.

Purpose of the Study:

• To develop a CV-oriented adaptive CS framework using saliency detection.
• To enhance the preservation of information critical for CV tasks.
• To optimize sensor data utilization and improve inference accuracy in CV applications.

Main Methods:

• Implemented a saliency detection algorithm to identify important image regions.
• Developed an adaptive CS framework prioritizing data relevant to CV tasks.
• Conducted experiments on real-world datasets (STL10, Intel, Imagenette, KITTI) using real sensor devices.

Main Results:

• Achieved significant improvements in classification accuracy (up to 26.23%) on benchmark datasets.
• Demonstrated superior performance in object detection on the KITTI dataset.
• Showcased robustness at low sampling rates compared to state-of-the-art CS techniques.

Conclusions:

• The proposed CV-oriented adaptive CS framework effectively prioritizes critical information for CV tasks.
• This approach enhances data acquisition efficiency and accuracy in sensor-rich IoT environments.
• The method offers a robust solution for reducing transmission costs while maintaining high CV performance.