DeblurGAN-v2-Based Landslide Recognition Method for Blurred Images

Abstract

Aiming at the problem of motion-blurred images due to camera shake, which affects the performance of landslide recognition when using computer vision technology for landslide recognition, a fuzzy image landslide recognition method combining DeblurGAN-v2 and YOLOv5 is proposed. Firstly, the custom camera image landslide dataset is trained by YOLOv5 deep learning object detection algorithm and a landslide recognition model is established; then the DeblurGAN-v2 algorithm is used to deblur the blurred image and obtain the deblurred image; finally, the deblurred image is inputted into the landslide recognition model to identify and locate the landslide area in the image. The experimental results show that the accuracy of landslide recognition of blurred image is 35%, and the accuracy of landslide recognition of deblurred image is 82.5%. Meanwhile, the statistical landslide confidence is found that the mean and variance of confidence are 0.26 and 0.13 under blurred image, and 0.67 and 0.11 under deblurred image, respectively, and the proposed method is able to effectively improve the accuracy of blurred image landslide recognition and the confidence. confidence, and has better stability, which provides an important technical solution for real-time monitoring, early warning and rescue decision-making of landslide disasters.