Feature localization based on microscopic vision is important for precision assembly. Because assembly states vary in a batch assembly， feature positioning errors often arise， which significantly interrupt the process and affect efficiency. Therefore， establishing a solid and robust feature localization algorithm is crucial. This paper proposes a support vector machine （SVM） model for synthesizing gradient histograms and local binary patterns. Furthermore， the pyramid search strategy is employed to improve the recognition efficiency and realize the micro-feature localization method. Performance verification and heuristic application are conducted on self-developed precision automatic assembly equipment， and different features are collected for SVM training. The influences of interference factors such as texture and illumination on the positioning stability are investigated in detail. Additional experiments regarding the positioning accuracy and actuator component assembly are performed. Under various conditions， the proposed approach presents good unimodal， repetitive accuracy and robustness. A recognition accuracy rate of 98% can be achieved. The positioning accuracy is better than 4 μm， and the actual assembly accuracy is better than 7 μm. The feature localization method can meet the localization requirements under different assembly conditions in real batch production and provides an effective solution for precision automatic assembly localization.