Microfeature positioning based on machine vision is a crucial aspect of precision automated assembly. External interference and differences in the parts themselves can easily cause visual guidance errors and the success rate of assembly. Therefore， a composite positioning method consisting of rough positioning and fine positioning was proposed. First， the region of interest was extracted through a target-frame-detection algorithm based on a convolutional neural network to achieve rough positioning. Based on this， precise positioning of parts was achieved through contour geometric feature registration. A dynamic learning mechanism assisted by automatic labeling was also adopted in the algorithm to solve the problem of the high positioning failure rate resulting from the difference between the different batches of parts. The method was tested on assembly equipment developed by the research group. The effects of brightness， defocusing， and posture changes on the robustness of visual positioning algorithms were analyzed. Furthermore， positioning accuracy and small-batch assembly experiments were conducted. The results show that the proposed method has good robustness and repeatability with various forms of interference， with an assembly success rate of 97%. Both the absolute accuracy and repetitive accuracy of visual positioning are<2 μm， and assembly accuracy is<10 μm. Therefore， the research results effectively meet the dual requirements of both accuracy and robustness of the positioning algorithm in precision microassembly.