First Gradually, Then Suddenly: Understanding the Impact of Image Compression on Object Detection Using Deep Learning

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Video surveillance systems generate vast amounts of image data.
• Lossy compression, like Joint Photographic Expert Group (JPEG), is used for efficient storage and transfer.
• However, JPEG compression can degrade image quality, potentially affecting subsequent analysis.

Purpose of the Study:

• To investigate the impact of JPEG image compression on the performance of object detection models.
• To analyze how varying compression levels affect the accuracy and robustness of deep learning-based object detectors.
• To provide guidance for practitioners on selecting appropriate compression levels for surveillance applications.

Main Methods:

• The study focused on Joint Photographic Expert Group (JPEG) compression.
• Performance metrics were thoroughly analyzed across a range of compression characteristics.
• Nine popular object-detection deep learning models were evaluated on a standard benchmark dataset.

Main Results:

• The research experimentally assessed the robustness of nine object-detection deep models against different JPEG compression levels.
• Significant variations in model performance were observed based on the degree of image compression.
• The study identified a trade-off between compression ratio and object detection accuracy.

Conclusions:

• Practitioners can use the findings to determine acceptable JPEG compression levels for specific video surveillance use cases.
• This methodology aids in managing large image datasets by optimizing compression without critically compromising object detection.
• The research contributes to efficient processing and long-term retention of surveillance imagery.