Distributed Learning and Inference With Compressed Images.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Modern computer vision systems process large datasets, often with resource constraints on data storage and transmission.
• Scenarios involving physically separated data capture and processing (e.g., autonomous vehicles, cloud computing) are common.
• Lossy image compression is vital for managing data under these constraints but can degrade image quality.

Purpose of the Study:

• To analyze the impact of lossy image compression on downstream computer vision tasks, focusing on semantic information loss and covariate shift.
• To propose a novel method for mitigating the negative effects of image compression on model performance.
• To address challenges in scenarios where training and inference data distributions differ due to compression.

Main Methods:

• Dataset restoration using generative adversarial networks (GANs) for image restoration.
• Analysis of performance degradation in computer vision tasks under varying compression rates.
• Focus on semantic segmentation as a use case for evaluating the proposed method.

Main Results:

• Lossy image compression leads to semantic information loss and covariate shift, degrading downstream task performance.
• The proposed dataset restoration method effectively alleviates the negative impacts of compression.
• Performance improvements were observed across diverse datasets and a wide range of compression rates.

Conclusions:

• Image compression poses significant challenges for computer vision, particularly in resource-constrained and distributed systems.
• Generative adversarial network-based dataset restoration offers a robust and adaptable solution.
• The proposed method enhances the reliability of computer vision models operating on compressed data without increasing deployment costs.