散斑法是表面粗糙度测量领域的研究热点之一，该方法可以通过建立散斑图像特征参数与表面粗糙度评定参数之间的关系，实现对工件表面粗糙度的高效和无损测量。然而，该方法在特征参数选取阶段缺乏统一的标准，工件的机加工方法也会对特征参数和表面粗糙度评定参数之间的关系产生影响。这可能导致选取的特征参数仅适用于某种加工工艺下的表面粗糙度测量，并且特征参数之间还可能存在冗余问题。针对以上问题，文中从采集的激光散斑图像中提取了多个特征参数，引入斯皮尔曼相关系数，制定简约规则对提取的特征参数进行预筛选，提出了改进的序列后向选择算法以剔除冗余特征。实验结果表明：文中提出的方法筛选出了一组与不同加工工艺的表面粗糙度均强相关的特征，并解决了特征冗余问题，利用这组特征建立的表面粗糙度测量模型能100%识别试件的加工类型，并实现对其表面粗糙度较高精度的测量，改进的序列后向选择算法将平磨、卧铣、立铣和研磨试件表面粗糙度测量模型的平均绝对百分比误差分别降低了1.22%、0.62%、4.99%和1.61%，解决特征冗余问题的同时建立的模型性能更优。

Speckle method stands out as one research hotspots in the realm of surface roughness measurement, boasting advantages such as low loss, high-temperature resistance, and high reliability. As the scenarios for surface roughness measurement grow in complexity and precision requirements continue to escalate, a novel surface roughness modeling method based on multiple feature parameters of laser speckle images has been proposed. This method is grounded in multiple feature parameters extracted from laser speckle images. However, the modeling process using this approach confronts challenges related to feature correlation and redundancy. The presence of irrelevant or redundant features in the modeling process can result in prelonged feature extraction times, heightened computational costs, and increased model complexity. Furthermore, these features can detrimentally affect the accuracy and stability of the model. To address these issues, a method is proposed to alleviate feature correlation and redundancy during surface roughness modeling. Simultaneously, the selected features are designed to facilitate the measurement of surface roughness across various processing types.Introducing Spearman's correlation coefficient, we aim to establish succinct rules for the effective screening of laser speckle image feature parameters that exhibit strong correlations with the surface roughness evaluation parameter R_a for each processing type (Tab.1). To address redundancy among feature parameters, an enhanced sequential backward selection algorithm is employed. Subsequently, laser speckle images from various peocessing types, including plane grinding, horizontal milling, vertical milling, and grinding polishing standard specimens, were acquired through experiments (Fig.2, Fig.3). Utilizing these collected laser speckle images, we constructed a surface roughness measurement model based on support vector machines. The method's efficacy was then validated through comprehensive verification processes. From the gathered later speckle images, a total of 27 feature parameters were initially extracted. By introducing Spearman's correlation coefficient and formulating simple rules, we identified 8 feature parameters {E, S, I, H, B_ent, κ, σ, υ} strongly correlated with the surface roughness evaluation parameter R_a for each processing type. Then, redundant feature parameters H, κ and σ were effectively eliminated using an improved sequence backward selection algorithm. This process not only addressed the issues of feature correlation and redundancy but also led to the establishment of a surface roughness measurement model incorporating the selected feature parameters{E, S, I, B_ent, υ}. The resulting model demonstrated a remarkable 100% recognition rate for processing type and exhibited high-precision measurement of surface roughness (Tab.7, Tab.8). In addition, the enhanced sequential backward selection algorithm contributed to a reduction in the MAPE for the surface roughness measurement model across different specimens: plane grinding, horizontal milling, vertical milling, and grinding polishing. The reductions were 1.22%, 0.62%, 4.99% and 1.61%, respectively. The proposed method effectively addresses the issues of feature correlation and redundancy in the process of surface roughness modeling, leveraging multiple feature parameters extracted from laser speckle images. By eliminating irrelevant and redundant features, the method prevents unnecessary consumption of feature extraction time and reduces model calculation costs. The resulting model exhibits enhanced stability and accuracy. Experimental results demonstrate the effectiveness of the model established using the selected feature parameters. It achieves a 100% recognition rate for processing types such as plane grinding, horizontal milling, vertical milling, and grinding polishing specimens. Moreover, the MAPE for surface roughness prediction is reduced to 3.55%, 3.10%, 3.17%, and 2.27%, respectively. These reductions represent improvements of 1.22%, 0.62%, 4.99%, and 1.61% compared to the model's perfomance before removing redundant features.