Glass microlens arrays, with calibration, focusing, diffusion illumination or imaging functions, are widely used in LED, OLED optics and artificial compound eyes. Hot embossing technology is a promising way for achieving mass production of glass microlens arrays. However, the glass material undergoes thermal expansion and mechanical deformation during hot embossing, which inevitably generates form errors in both macro and micro scales, thus deteriorating its optical performance. Therefore, there is a need to implement efficient and reliable quality control of high-volume glass microlens arrays produced hot embossing. Measurement instruments such as Taylor Hobson probe profilers and white light interferometers are of high-cost and low-efficiency, making it unsuitable to be used for inspection of high-volume and large-area glass microlens arrays. As a result, this study developed a machine-vision based measurement system for inspection of macro and micro features of glass microlens arrays formed by hot embossing. The highlight is that the machine-vision based inspection method with a modified image stitching algorithm is applied in efficiently evaluating the replication deviation of thermally embossed glass microlens array features for the first time.Figure 3 shows the home-made vision inspection device for micro-optical components, which consists of two modules: macro inspection and micro inspection. The macro detection module consists of a 6-million-pixel CMOS camera, a lens with a low magnification of 0.7×, and a backlight. The microscopic inspection module consists of an 8-million-pixel CMOS camera, a lens with a high magnification of 40×, coaxial light source, and a XY linear stage with an accuracy of ±1 μm. In microscopic measurements, the precision positioning is realized by the XY linear stage and the linear encoders. Fig.4a shows the flowchart of the macro detection. First, the checkerboard pictures were taken by the CMOS camera. The internal and external parameter matrix and distortion parameters of the camera were obtained, and the system calibration was carried out to obtain the pixel equivalent, followed by the correction of the images of the glass optical components. After image graying and median filtering, the Canny algorithm was used for edge detection. The next step is the improved Zernike moment subpixel edge extraction. After that, the least squares method was adopted to fit the edge coordinate points of the optical component. Finally, the diameter of the glass optical component was calculated. Figure 4(b) shows the flowchart for the microscopic detection. The sapphire mold and the hot-embossed glass microlens array were clamped on the XY linear stage. The linear stage and CMOS camera were controlled to acquire images of the microlens array on the surface of the mold and glass replica during the movement. The acquired images are automatically searched for the best template to match image features for implementing image stitching. After the rotation treatment, the images of the microlens array in corresponding positions of the mold and the glass replica were intercepted. Next, the area of the microlens contour was filled. The Canny algorithm was then used for edge detection of the microlenses. After that, the least-square method was adopted to fit the edge coordinate points of the microlens array at the corresponding position of the mold and the glass replica. Finally, the geometric features such as diameter, center coordinates, and roundness of microlenses on the mold and the glass replica were calculated. Moreover, the replication deviation of the embossing glass microlens array was calculated.Taking the measurement results of the spiral micrometer as the real value, the absolute error and relative error of the measurement results of the machine vision are calculated and shown in Tab.1. It is seen that the average error of the improved Zernike moment subpixel algorithm is 3.7 μm in measurement of the diameter of glass optical components. Table 3 compares the measurement results of the self-developed vision inspection device and the white light interferometer at different magnifications. The measurement results of the self-developed device are closest to those of the white light interferometer magnified by 50×. Moreover, stitching of 15 images can be completed within 20 s by using the method of searching the best template matching, which proves the efficiency and accuracy of the inspection device. The experimental results show that the diameter of the optical component increases by 15.6 μm after hot embossing. Replication deviation map is established for a selected area of about 850 μm×150 μm, so that the center coordinate deviation of the microlens at the corresponding positions of the glass optical component and the mold can be visually displayed. It is found that the center distance between the two microlenses on the diagonal of the glass optical component is 1.423 μm higher than that on the mold.To meet the demand for quality inspection of high-volume hot embossed glass microlens arrays, a machine-vision based measurement system was developed for measuring macro and micro feature sizes of glass microlens arrays formed by hot embossing for the first time. A circular glass microlens array sample with a macro diameter of ~7.44 mm and a microlens diameter of ~10 μm was subjected to the measurement validation test. The experimental results show that the diameter of the optical element increases by 15.6 μm after hot embossing and the average error of the measured diameter is 3.7 μm. The replication deviation map was established for a selected area of about 850 μm×150 μm. It is found that the center distance between the two microlenses on the diagonal of the glass optical component is 1.423 μm greater than that on the mold. By comparing with the measurement results of the spiral micrometer and the white light interferometer, the efficiency and accuracy of the self-developed machine-vision based inspection device are demonstrated. Furthermore, this measurement system is also suitable for the deviation analysis of microlens arrays produced by other replication technologies beyond hot embossing.