Abstract: Surface defect detection for PCB, the speed and accuracy of detection need to be improved, An image detection algorithm: FFS-YOLO is proposed, which is based on the YOLO-V4-tiny framework. The calculation process of the algorithm is: Firstly, the optimized K-means clustering method was used to cluster the defect data set to solve the problem that the initial prior frame was not suitable for PCB surface defect detection. Secondly, in order to solve the problem of small-scale target information loss during down sampling, FOCUS slicing operation is introduced. Thirdly, SCC structure was introduced into PANet to improve the model receptive field and enhance the expression ability of small target features, so as to optimize the model performance. Finally, the Focal loss is used to optimize the loss function. The dataset used for algorithm experimental verification comes from the PCB surface defect dataset published by Peking University, the results showed that: the average detection accuracy of FFS-YOLO was 99.22%, FPS was 142, and the number of model parameters was 6.10MB. Compared with the classical algorithm, the detection speed, accuracy and the number of model parameters of FFS-YOLO are greatly improved.

Key words: PCB surface defect detection, YOLOv4-tiny, K-means, Focus slice, Focal?Loss