Micro-Electro-Mechanical Systems (MEMS) have the characteristics of miniaturization and high integration. As the high aspect ratio of MEMS increases, the measurement of MEMS feature size faces greater challenges. Through-focus Scanning Optical Microscopy (TSOM) technology is a high-precision and nondestructive optical measurement method. TSOM images are captured along the scanning direction by collecting a set of defocused images and the size information of the structure is extracted from TSOM images by the library matching method. This method is highly sensitive and suitable for nano-scale structure measurements, but it is difficult to build a database for micron-scale features and is susceptible to environmental interference. In this paper, a TSOM optical system is established and traditional optical microscopy is used to collect a set of defocused images. The TSOM’s feature vector set is obtained by the image feature extraction method and is combined with machine learning to establish MEMS groove regression prediction models with different feature sizes. The results show that the above method can achieve nano-scale high precision measurement of a MEMS groove width and the single point repeatability measurement has great performance. The Relative Standard Deviation (RSD) of 2 μm width is about 1%, and the RSD of 10 μm and 30 μm width are respectively lower than 0.2% and 0.35%. This method has very high application prospects for micron MEMS groove structure measurement.