金属发动机叶片微小缺陷超分辨图像重建方法

Super-Resolution Image Reconstruction Method for Micro Defects of Metal Engine Blades

针对金属发动机叶片细微缺陷边界对比度低、描述符不足等造成的检测困难问题,提出一种基于超分辨图像重建技术实现微小缺陷强化的方法。首先,为弥补固定分辨率常规图像量化能力不足的短板,设计基于光度立体的图像重建方法,在图像量化层面实现了叶片表面高对比度法向贴图重建;然后,针对微小缺陷采样描述符不足的问题,通过真实叶片图像来构建多角度、多姿态的数据集,采用基于像素损失的Charbonnier损失来改进超分模型的损失函数,从而实现图像的高分辨率重建,强化采样分辨率,最终实现量化以及采集两个层面的微小缺陷超分辨增强;最后,使用传统的Canny算子识别叶片表面缺陷边界。实验结果表明,所提方法可以免疫二维歧义性,相较于传统方法,最高可提升金属叶片表面微小缺陷检出率达24.3%。

Objective Aiming at the difficulty of detection caused by low contrast and insufficient descriptors of micro defects in metal engine blades,this paper proposes a super-resolution image reconstruction technique to enhance micro defects.Various kinds of tiny defects may occur during the manufacturing and use process of metal aero-engine blades,which will have a huge impact on the appearance of the product or even the overall function.Therefore,the detection of tiny defects on the metal surface has profound significance for the overall product quality control and loss assessment of parts.Current detection methods are mostly based on manual detection,which has low reliability.The main factors that make it difficult to identify defects accurately are unclear feature boundaries and low contrast between defect contours and background,other noise in images or the two-dimensional ambiguity interference,and tiny defects with insufficient image descriptors for accurate identification.To address the above problems,researchers have proposed corresponding solutions from the perspective of image enhancement and fusion reconstruction.However,both image enhancement and image fusion methods start from the overall image information,such as adjusting the histogram,contrast,and other comprehensive attributes of the image to strengthen the features of the target,which are prone to problems such as negative optimization and two-dimensional ambiguity interference.Therefore,this paper performs image enhancement from the imaging principle and designs the image feature enhancement technique from the quantization and sampling aspects of image digitization respectively.Methods The image digitization includes two processes:sampling and quantization.With 8 bit grayscale images as examples,the discretization of the continuous coordinates of the image space is called sampling,and the grayscale values of some points,also called sampling points,in the space represent the image.The conversion of the grayscale values of the sampled pixels from analog to discrete quantities is called the quantization of the image grayscale,which determines the gray-level resolution of the image.The super-resolution quantization sampling enhancement technique for images of tiny defects on metal surfaces mainly focuses on contrast enhancement,resolution enhancement,elimination of twodimensional illumination unevenness,and stain effects of tiny feature details of images.It can reveal low-contrast and border-unclear details in a way that can be more easily recognized by human eyes and computers while retaining the original clear features of images.In this paper,image enhancement is performed from the principle of imaging technology,and a two-dimensional super-resolution enhancement technique with fused image acquisition and quantization is designed.As the photometric stereo has the characteristics of refined normal mapping reconstruction,this paper proposes an image enhancement reconstruction technique based on photometric stereo and image hyper-segmentation to address the problems of existing methods.For the shortcomings of the quantization level in the digitization process,it uses photometric stereo technology for the high-contrast display to highlight the image contour features,overcoming the deficiency in the previous image with low contrast of fine features and vulnerability to two-dimensional ambiguity interference.For the deficiency of sampling resolution level in the image digitization process,the image hyper-segmentation reconstruction method is introduced to solve the problems of insufficient details and discrete image descriptors caused by the hardware bottleneck in the traditional photometric stereo technology.Results and Discussions Firstly,to compensate for the shortage of quantization capability of fixed-resolution conventional images,the paper designs a photometric stereo-based image reconstruction method to achieve high-contrast normal mapping reconstruction of the blade surface at the image quantization level(Fig.11).In addition,in terms of insufficient sampling descriptors for small defects,the multi-angle and multi-pose dataset is constructed from real blade images,and the loss function of the super-resolution model is improved by Charbonnier loss based on pixel loss.Additionally,the appropriate hyper-parameters are configured to reconstruct the high resolution and enhance the sampling resolution.The resolution-enhanced images can improve the original image pixel count to two to four times the original(Fig.15),enhancing image details.Eventually,the super-resolution of small defects at both quantization and acquisition levels is enhanced(Fig.19).The proposed enhanced reconstruction method finally uses the traditional Canny operator to identify the defect boundary of the blade surface.The experimental results show that the proposed method is immune to two-dimensional ambiguity and can improve the detection rate of minor defects on the metal blade surface by up to 24.3%(Table 5)compared to the traditional method.Conclusions The application of the image super-resolution reconstruction technique proposed in this paper can effectively improve the recognition rate of metal blade surface defects and reduce the false detection rate caused by two-dimensional ambiguity.The experimental results show that the recognition rate of minor defects on the metal blade surface can be improved by 24.3%compared with the traditional method.Especially in the case of stains on the blade surface and poor lighting effect of grayscale images,the image quantization contrast enhancement can shield the non-defective features such as stains and strengthen the display contrast of the surface,and the image sampling information enhancement can be pixel intensive and reduce the defects ignored due to too few pixels.The proposed method has a good prospect of application in industrial static inspection.Compared with the existing methods,the proposed method is applied in the image input preprocessing stage and can be easily integrated before the defect detection operator,which is conducive to promotion and popularization.Subsequently,it is possible to extend the applicability and improve the robustness of image fusion reconstruction with two-dimensional information enhancement by utilizing a streamlined and lightweight network to conduct targeted data training on detection objects and integrating the hardware structures of photometric stereo.