Hardness is a performance index to measure the degree of softness and hardness of materials. In view of the characteristics of lutetium-aluminum garnet (LuAG), lutetium-gallium garnet (LuGG) and other optical crystals with small elastic compression, easy to break and unclear indentation boundary. An optical crystal Vickers hardness measurement method based on image processing was proposed. A deep learning YOLOv5s network was used to segment the image. The exact positions of the four vertices were obtained by adaptive binarization, maximum connected domain selection, skeleton extraction and then specific directional line segments detected by probabilistic Hough line treatment. The experimental results show that the average relative error can be controlled within 1.5%, which effectively reduces the calculation error of traditional network algorithm and traditional image processing algorithm, and is suitable for automatic and accurate measurement of optical crystal Vickers hardness.