结构光熔池传感中反射图的递归选区处理算法

Recursive-Selective Processing Algorithm of a Reflection Image in Structured Light Weld Pool Sensing

点阵结构光三维熔池传感中处理反射图并识别成像点是后续熔池重构计算的基础,然而,由于弧光在成像屏上分布不均匀,加上结构激光照射到工件表面后漫反射到成像屏上,生成了额外的高亮背景,因此大大增加了成像点识别难度.此外,激光束在传播过程中受金属蒸气散射发生横向扩散,造成成像屏上的成像点尺寸变大、对比度下降,给识别带来了更大困难,极易丢失成像点(丢点)或误将噪声识别为成像点(多点).针对该问题,提出了一种递归选区图像处理算法,该算法由整体到局部,利用递归的思想不断选择“未成功识别”区域做进一步处理,当逐层返回处理结果后,可实现从不均匀背景中分离出所有成像点,不易丢点或多点.在每层处理计算中,算法的主要步骤包括阈值、滤波、连通域计算、大连通域递归处理以及小连通域重新覆盖等.阈值处理采用OTSU算法,该算法针对各层目标图像的亮度特征自动确定最佳阈值.滤波处理采用中值滤波法,提出每深入两层将滤波窗口的尺寸减小2个像素,可以降低丢点可能性,减少无用递归,避免超过最大递归深度.最后,用起弧后不同时刻拍摄的具有不同特征的反射图验证了递归选区处理算法的有效性,并分析讨论了该算法的实时性能.结果显示,单个图像的处理平均用时约为46 ms,可以满足实时传感要求.

Processing reflection images and identifying imaging points in structured light sensing are the bases of subsequent reconstruction of a 3D weld pool surface.However,due to uneven distribution of arc light on the imaging plane and diffuse reflection of structured laser light projected onto the workpiece surface,identification could not be achieved effectively.Moreover,laser beams get scattered by metal vapor during propagation,which enlarged the sizes of some imaging points and decreased their contrast with the background,adding to difficulties in identification easily missing imaging points or mistakenly taking noises as imaging points.To solve this problem,a recursiveselective image processing algorithm was proposed.From whole to local,the algorithm used the idea of recursion to continuously select those unidentified areas for further processing.When the processed areas were returned layer by layer,all imaging points could be separated from the uneven background,with no missing or redundant points.At each layer,the main steps of the algorithm included thresholding,filtering,computing connected domains,recursive processing of large connected domains,and recovering of small,connected domains.The OTSU algorithm was used for thresholding,which automatically determined the best threshold of the target image at each layer according to its brightness characteristics.Median filtering was adopted,and it was proposed to reduce the size of the filter window by 2 pixels every two layers,which would reduce the possibility of missing points and avoid invalid recursion in case the maximum recursion depth was exceeded.Finally,reflection images with unique characteristics taken at different moments after arcing were used to verify the effectiveness of the proposed algorithm,and real-time performance of the algorithm was also analyzed.The results show that the average processing time of one image is about 46 ms,which meets the requirements of real-time sensing.