低分辨率图像目标识别的方法、装置及系统
By combining a super-resolution image generator and a high-resolution image classifier, and using Gaussian blur data augmentation and label smoothing loss function training, the problem of low target recognition accuracy in low-resolution images is solved, achieving higher recognition accuracy and better classifier learning effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ARMY ENG UNIV OF PLA
- Filing Date
- 2023-03-10
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies struggle to effectively improve recognition accuracy in low-resolution image target recognition, especially in non-face and traffic sign recognition tasks. Furthermore, existing methods often require the construction of two branch networks, one for high resolution and one for low resolution, making it difficult to expand applications.
A super-resolution image generator is used to generate super-resolution images, and a high-resolution image classifier is used for recognition. The high-resolution image classifier is trained by Gaussian blur data augmentation and label smoothing loss function, and the super-resolution image generator is trained by combining super-resolution loss and perceptual loss to avoid contradictions and conflicts between the generator and the classifier.
It significantly improved the recognition accuracy of targets in low-resolution images. Experimental results showed an improvement of 2% to 6% in recognition accuracy, enhancing the learning effect and recognition ability of the classifier.
Smart Images
Figure CN116452418B_ABST