基于硬件的复杂光照环境多材质机器视觉特征提取研究

Hardware-Based Research on Machine Vision Feature Extraction for Multi-Material Targets in Complex Illumination Environments

摘要: 在智能制造非受控复杂光照场景中，常规可见光机器视觉系统易受强杂散光干扰，产生高光过曝、阴影畸变等问题，造成伪边缘误检与几何细节丢失，而传统后端算法补偿存在时间延迟、无法还原感光超限丢失信息的固有瓶颈。对此，本文将抗干扰环节物理前置，提出一套“近红外主动照明 + 窄带光谱滤波 + 微秒级硬件同步触发”的软硬件协同抗干扰成像方法。该方法基于太阳光谱940 nm水汽吸收谷的天然物理特性确立工作波段，搭建了配套的成像与照明硬件平台，建立多维图像质量客观评价体系完成高反光、强吸光、漫反射三类典型工业材质的成像参数自适应寻优，并通过Sobel与Canny经典算子验证边缘提取效能。实验结果表明，该方法可从成像源头有效抑制杂散光干扰，显著降低背景灰度波动，将三类工业材质的伪边缘像素占比压缩至0.01%以下，大幅提升了复杂光照下多材质目标边缘特征提取的精度与鲁棒性，拓宽了机器视觉在非受控工况下的应用边界。

Abstract: In uncontrolled complex illumination scenarios of intelligent manufacturing, conventional visible light machine vision systems are susceptible to strong stray light interference, which causes problems including highlight overexposure and shadow distortion, further leading to false edge misdetection and loss of geometric details. Traditional back-end algorithm compensation, however, has inherent bottlenecks: it introduces time delay and is unable to restore information lost due to sensor saturation. To address this challenge, this paper physically advances the anti-interference link to the front end of the imaging chain, and proposes a hardware-software collaborative anti-interference imaging scheme integrating near-infrared active illumination, narrow-band spectral filtering, and microsecond-level hardware synchronous triggering. Based on the natural physical characteristics of the water vapor absorption valley at 940 nm in the solar spectrum, this method determines the optimal operating waveband, constructs a matched imaging and illumination hardware platform, establishes a multi-dimensional objective evaluation system for image quality to complete adaptive optimization of imaging parameters for three typical industrial materials (high-reflective, strong light-absorbing and diffuse reflective materials), and verifies its edge extraction performance via the classical Sobel and Canny operators. The experimental results show that the proposed method can effectively suppress stray light interference from the imaging source, significantly reduce background gray level fluctuation, compress the proportion of false edge pixels for all three types of industrial materials to less than 0.01%, greatly improve the accuracy and robustness of edge feature extraction for multi-material targets under complex illumination, and broaden the application boundary of machine vision in uncontrolled working conditions.