【摘要】 高动态范围物体表面缺陷会严重影响其性能与可靠性。为实现高效率、高精度的在线检测，提出了一种基于调制照明的高动态范围物体表面在线缺陷检测方法。该方法通过三角波调制照明光源对运动中的高动态范围物体表面进行周期照明，相机以固定帧率采集多光照条件下的图像序列，采用基于差分统计修正的特征对齐方法实现图像配准，并通过自适应权重的多曝光融合算法重建高动态范围检测图像。实验结果表明，该方法可有效检测微透镜阵列和玻璃-金属复合结构等不同反射率表面的缺陷，适用于大批量生产制造和在线检测等场景。更多还原

【Abstract】 Objective Objects with high dynamic range(HDR) surfaces—including highly reflective metals and alloys, optical components(such as micro-lenses and diffraction gratings), semiconductor wafers, heterostructure devices, and biological tissues, have significant application value in precision manufacturing, metrology, process engineering, energy systems, semiconductors, and medical diagnostics. In lithium-ion battery production, electrode sheet surface defects can severely compromise battery performance and safety, subsurface lens defects degrade optical properties, and wafer surface defects directly reduce chip yields. Thus, efficient and precise HDR surface defect identification is critically important. Traditional manual visual detection suffers from inefficiency, high labor intensity, and inconsistent quality, failing to meet large-scale detection requirements. Although machine vision improves defect detection efficiency and accuracy, conventional cameras struggle with HDR surface imaging under standard illumination, facing issues like specular highlights and nonuniform reflectivity. Specifically, high reflectivity causes image saturation, while low reflectivity leads to defect omission. Achieving efficient, high-precision online defect detection for HDR surfaces remains a key challenge in optical defect detection.Methods This paper proposes an HDR online defect detection method based on modulated lighting to enable high-speed, highprecision defect detection on HDR surfaces. Traditional multi-exposure acquisition methods require multiple captures at the same position by adjusting the camera’s exposure time, which reduces the frame rate, failing to meet high-speed online detection requirements. In contrast, the proposed triangular-wave-modulated lighting technique allows the camera to maintain a high frame rate continuously, significantly improving detection efficiency. Notably, object motion and environmental interference(such as mechanical vibration, camera shake, and stage speed fluctuations) induce non-uniform displacement in the multi-illumination image sequence. To address this issue, we first employ the speeded-up robust feature(SURF) algorithm for image alignment. However, since insufficient feature point extraction under extreme lighting(low-illumination or high-illumination conditions) degrades alignment accuracy, we further propose a differential statistical correction-based feature alignment method for feature alignment. Finally, HDR image reconstruction is performed using an adaptive weighted multi-exposure fusion algorithm.Results and Discussions The post-processing of multi-illumination images begins with feature alignment to register the acquired image sequences. Due to relative displacement between the original images, the middle image is selected as the reference frame, and feature matching points are computed for the remaining images relative to this reference. The distribution of feature matching points follows a Gaussian distribution with respect to illumination intensity: the highest density of matches occurs near the reference frame, while the number of matching points decreases significantly in low-illumination and high-illumination images at the extremes. This reduction in feature matches leads to increased registration errors in under-and overexposed images. Theoretically, the displacement between images should follow a linear relationship; however, empirical observations reveal significant deviations at both illumination extremes. To mitigate this issue, we propose a differential statistical correction-based feature alignment method, which corrects displacement outliers using the mean displacement of normally exposed images. After correction, the displacements conform to a linear relationship. Next, an adaptive weighted multi-exposure fusion algorithm reconstructs the aligned image sequence into an HDR image. The test sample contains two areas with distinct surface reflectivities, each exhibiting different defect types: linear defects in high-reflectivity areas and point defects in low-reflectivity areas. In low-illumination images, linear defects are clearly visible in highreflectivity areas, while point defects in low-reflectivity areas are obscured due to inadequate brightness. Conversely, in highillumination images, overexposure masks linear defects, but point defects become discernible. In contrast, the HDR images reconstructed by our method resolve both defect types clearly across all reflectivity conditions.Conclusions This study proposes an HDR online defect detection method using modulated lighting to address the online challenge of inspecting HDR objects. The system employs an active modulated lighting setup with a fixed-frame-rate camera to acquire image sequences under varying illumination conditions. A differential statistical correction-based feature alignment method ensures precise image registration across lighting variations. Furthermore, an adaptive weighted multi-exposure fusion algorithm reconstructs HDR detection images. Validation experiments on microlens arrays and glass-metal composite structures demonstrate the method’s capability to simultaneously detect defects on surfaces with varying reflectivity. Results reveal that camera frame rate limitations induce motion blur during high-speed detections of HDR objects, compromising detection efficiency. Future work will investigate high-frame-rate cameras or frame rate optimization strategies to enhance system performance.更多还原