Nowadays, the mainstream method of LED defects detection is low-efficiency manual visual inspection. And the traditional machine visual detection cannot meet its application’s standard with its low precision. To deal with these issues, a LED defects detection method is proposed based on improved faster R-CNN network framework. In order to improve the robustness and generalization capability of the network, the dataset is expanded by adding noise and changing the brightness. Resnet50 and FPN network are the backbone network to extract the characteristics, and the anchor scale is adjusted according to the characteristics of its different feature prediction layers of the feature pyramid, to construct and train the network. Eventually, the quantitative analysis of the test results on the dataset testing shows that the method of LED bubble-like defects detection can achieve an overall accuracy of 95. 6%, and the recall rate has a 20. 8% increase. Single-picture detection time is about 100 ms. It can be affirmed that this method can meet the needs of the automatic detection in manufacture.