Aiming at the low detection accuracy caused by similar defect features, small defect target and large difference in defect scale in SOP chip defect detection, this paper proposes a defect detection method based on improved YOLOv8. The problem of information loss in the process of convolution pooling is solved by using SPD-Conv module. And introducing the SimAM attention mechanism, the model can learn the information in the 3D channel and improve the model′s perception of defect features. At the same time, BiFPN was used to replace the original feature extraction network, and multi-scale feature fusion was used to enable the model to better distinguish the defects with similar features and large-scale differences. Finally, a small target detection header is added to transmit more low-order feature information to the high-dimensional detection network to improve the detection effect of small target defects. Experimental data show that compared with the original model mAP@0.5/% increased by 5.4%, mAP@ 0.95/% increased by 4.3%, recall rate increased by 3%, has significant advantages compared with other models. In the generalization experiment, the mAP@0.5 of the improved algorithm is also improved by 2.7% compared with the original model, and a relevant system is designed to verify the effectiveness of the algorithm.