Edge Detection in Roof Images Using Transfer Learning in CNN

Abstract

Edge Detection in image processing is very important due to large number of applications it offers in variety of fields that extend from medical imaging to text and object detection, security, mapping of roads, real time traffic management, image inpainting, video surveillance and many more. Traditional methods for edge detection mostly rely on gradient filter based algorithms which usually require excessive pre-processing of the images for noise reduction and post-processing of the generated results in order to get fine edges. Moreover, traditional algorithms are not reliable generally because; as the noise in images increases their efficiency is affected largely due to escalation of mask size which also makes the system computationally expensive. In this paper, we will employ transfer learning in CNN method to detect edges of roof images. Incorporating CNN into edge detection problem makes the whole system simple, fast, and reliable. Moreover, with no more extra training requirement and without any additional feature extraction, CNN can process input images of any size. This technique employs feature map of the image using Visual Geometry Group (VGG) CNN network followed by application of Roberts, Prewitt, Scharr and Sobel edge operators separately to compute required edges. Interpretations of ground truths were obtained using manual techniques on roof images for performance comparison, and PNSR value of computed results via multiple operators against the ground truths is calculated.