Abstract: Printed circuit board assembly (PCBA) production is composed of a few complex processes, including tinning, placement, dispensing, reflow, and other production technologies, e.g., quality and function inspection, in which some processes require customized automation solutions. Moreover, industrial image detection has highly practical requirements in speed and accuracy. However, many theoretic algorithms can not meet the practical requirements. Therefore, it is necessary to select appropriate and effective detection algorithms according to the actual characteristics of PCBA detection images. The objective of this paper is to process the images of non-standard automated test system with four-row mobile phone profiled (L-shaped) veneers in the RF automatic test line. To this end, this paper proposes an improved AKAZE feature detection algorithm to attain a strong robustness when the scale, rotation and illumination change and realize the low feature point description dimension and quick matching, and further overcome the limitation in the uniqueness and stability of conventional methods under abnormal conditions and actualize the accurate judgment of PCBA acupoint or fixture abnormality. Since the traditional gray classification method often suffers from batch classification errors, the support vector machine (SVM) is introduced to judge the station and fixture states. A SVM classification model is established to classify and train the characteristic sample library. The accuracy and timeliness of the training method are verified by using the local sample database until satisfying the application conditions of the existing solutions in the industrial environment, which shows that this proposed method can significantly improve the effect of simple classification method used in actual production, and reduce the occurrence of missed inspection and misjudgment. Finally, the cross-platform code migration is completed and the PCBA anomaly detection is applied to the real industrial environment. It is seen from the results that the proposed algorithm process can achieve the expected research objectives and meet the practical application requirements perfectly.