基于噪声估计的钢材表面图像增强与缺陷检测

Image enhancement and defect detection of steel surface based on noise estimation

低光照低对比度的钢材表面图像(low-light and low-contrast steel surface images,LCSI)往往被大量噪声污染,给检测和识别带来很大的困难,导致缺陷的识别率很低。为了解决这一问题,本文提出一种基于噪声水平估计(noise level estimation,NLE)的钢材表面图像分解增强算法。根据快速的噪声水平估计确定总变分(total variation,TV)正则化的平衡因子,将低光照低对比度的钢材表面图像分解成基础层和细节层,利用视网膜大脑皮层理论(retina+cortex,Retinex)模型将基础层分解为光照分量和反射分量并分别增强使光照均衡化。对于包含更多图像细节(缺陷)和噪声的细节层,使用高斯滤波抑制噪声后,再对细节进行增强并与增强后的基础层重构得到高质量的输出图像。最后利用最新的基于Canny边缘检测和基于大津算法(nobuyuki Otsu method,Otsu)对增强的钢材表面图像进行缺陷检测。实验结果表明:增强后的缺陷识别率比最新的方法提升至少15%以上。

Low-light and low-contrast steel surface images(LCSI)are often contaminated by lots of noise,which brings great difficulties to detection and recognition,resulting in a low defect-recognition rate.To solve this problem,this paper presents a decomposition and enhancement algorithm of steel surface image based on noise level estimation(NLE).Specifically,first,determining the balance factor of total variation(TV)regularization according to the fast NLE,and decomposing the LCSI into base and detail layers.Then,using the Retinex model to decompose the base layer into two components,i.e.,illumination and reflection,and enhancing the light of both to equalize state.For the detail layer that contains more image details and noise,Gaussian filtering is used to suppress the noise,and then the details are enhanced and reconstructed with the enhanced base-layer to obtain a high-quality output image.Finally,the latest Canny-based edge detection and Otsu-based algorithm are used to detect defects on the enhanced steel surface image.Experimental results show that by comparison with the latest method,the enhanced defect-recognition rate has increased by at least 15%.