The defects of ceramic substrates have a significant impact on the performance of electronic devices. To enhance the accuracy of defect detection, in this paper, based on the detection method of ceramic substrates by ultrasonic microscopy scanning, an improved neural network algorithm of YOLOv5 is proposed. Taking advantage of the penetrability of ultrasonic detection, a new backbone network is added to comprehensively integrate the echo information from both the surface and interior of the ceramic substrates. Meanwhile, a polarization attention mechanism is employed for feature fusion to improve the detection precision, and a lightweight network is integrated to reduce the number of parameters. Experiments of ultrasonic microscopy scanning on ceramic substrates were carried out to analyze the defect characteristics and create a dataset. On this dataset, the FusionPol-YOLOv5 model proposed in this paper achieves an average accuracy of 88.3% for the detection of 9 types of defects, with an mAP@0.5 of 91.7%. It can significantly reduce the human and material resources consumption and costs in the detection of ceramic substrates.