This publication describes techniques and processes for correcting distortion in an image in order to improve object detection by an object detector on an imaging device. In order to avoid missing target objects (e.g., faces) in an image during object detection due to distortion, the object detector performs object detection on the image using a low-threshold value. A low-threshold value is associated with a lesser chance of missing target objects. The detection results are compared against regions of the image that are known to have distortion due to factors, such as a wide field of view (WFOV) lens of the camera. The overlapping areas of the detection results and the distorted regions are identified as candidate areas that would benefit from distortion correction. Using an algorithm, the candidate regions are corrected (e.g., undistorted, cropped, down-sampled, rotated, and/or frontalized) to reduce distortion. Object detection is performed again over the corrected candidate regions, resulting in improved confidence. The object detection results can then be used by the imaging device to provide a high-quality image and a positive user experience with the imaging device.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Yang, Ruiduo; Shih, Yichang; Liang, Chia-Kai; and Hasinoff, Sam, "Improved Object Detection in an Image by Correcting Regions with Distortion", Technical Disclosure Commons, (April 01, 2020)