Research On Image De-roping Based On The Cloud

Object recognition technology has always been an important research direction in the field of computer vision,and it plays a key role in aerospace,medical,driverless,security monitoring and other fields.The Kuang-Chi Cloud is a high-altitude monitoring system that uses a cable rope to pull the airbag.Since the aerial photography system is suspended below the airbag,the captured image inevitably contains the rope information.These rope information seriously hinders the analysis of the scene by the information processing algorithm,so it needs to be eliminated before the scene analysis.The steps of rope culling in image mainly include two parts: the rope detection in image and image inpainting.The key is the rope detection in image.This paper first analyzes the feature extraction algorithm based on Hough transform,then we studies the feature extraction method based on edge contrast,a connected domain regression method based on linear regression is proposed to detect the connected domain of the rope,the method can detect most of the long length of the rope,but it is difficult to apply in practical scenes with complex scenes and curved knots.The rope detection algorithm based on deep learning is in the field of semantic segmentation,then this paper focuses on the semantic segmentation related algorithm of deep learning.because the proportion of the rope pixels is much lower than the ratio of the background pixels.This extreme situation will bring difficulties to the convergence of the model.Therefore,an improved model based on U-net is proposed,which introduces batch normalization to reduce over-fitting,using He initialization parameters to improve the stability of convergence,due to the proportion of the string pixels in the image is too small,a joint loss with cross entropy and Dice is proposed.The loss function is used to improve the detection accuracy and stability of convergence.In order to obtain more accurate segmentation results,OTSU is used for quadratic precise segmentation.The final model is able to quickly and accurately detect the rope in image withfewer samples in less time.The experimental results show that the Io U of this algorithm reaches 62.8% and the PA reaches 98.66%.The fast marching method algorithm is used to inpainting the rope image,and a very good result of the rope is obtained.The method of the image de-roping in this paper provides technical support for the practical application of The Cloud image system.