基于主成分分析的光伏热斑红外图像混合噪声去噪方法

为了解决光伏板热斑故障检测时受噪声影响的红外图像分辨率低而导致热斑区域难以识别的问题,提出一 种基于主成分分析的红外图像混合噪声自适应去噪方法。 该方法通过自适应窗口预处理算法将获取的热斑红外 图像进行初步去噪,滤除图像中的低密度椒盐噪声,减小噪声信号对后续选取降噪训练集时所造成的影响;然后, 采用基于块匹配的主成分分析法对预处理后的图像信息进行降维处理,提取信号的主要特征,降低噪声滤除时的 计算复杂度;最后,使用线性最小均方误差估计对图像进行二次去噪处理,滤除残余噪声;此外,在二次去噪之前重 新计算图像噪声水平,使最终的去噪图片获得了更好的视觉效果。 实验结果表明:该方法能够有效去除光伏热斑 红外图像中的混合噪声,客观评价指标显示噪声较小时,图像结构相似性可保持在 0. 9,在高密度噪声影响下,峰值 信噪比相较于修正的阿尔法均值滤波算法平均提高 2 dB,实际视觉效果中保留了图像细节特征,可以明显观测到 热斑区域。

Denoising Method of Mixed Noise in Photovoltaic Hot Spot Infrared Images Based on Principal Component Analysis

In order to solve the problem of difficult identification of hot spot areas due to the low resolution of infrared images affected by noise during hot spot detection of photovoltaic panels an adaptive denoising method of mixed noise in infrared images based on principal component analysis was proposed. The method used adaptive window pre-processing algorithms to initially denoise the acquired hot spot infrared images filtering out the low-density salt and pepper noise in the images and reducing the impact of the noise signal on the subsequent selection of the noise reduction training set. Then principal component analysis based on block matching was used to reduce the dimensionality of the pre-processed image information to extract the main features of the signal and reduce the computational complexity in noise filtering. Finally the image was denoised for the second time using the linear minimum mean square error estimation to filter out the residual noise. In addition the image noise level was recalculated before the secondary denoising so that the final denoised image obtained a better visual effect. The experimental results showed that this method effectively removed the mixed noise in the photovoltaic hot spot infrared image. The objective evaluation index showed that the structural similarity of the image was maintained at 0. 9 when the noise was low. The peak signal-to-noise ratio of the image under the influence of high-density noise was increased by 2 dB on average compared with the modified Alpha mean filter algorithm. The image details were retained in the actual visual effect and the hot spot area could be observed significantly.