Aiming at the problems of poor effect and inaccurate defect location of traditional steel plate surface defect detection methods, a deep learning detection algorithm based on improved RetinaNet-GHM was proposed. Firstly, the path aggregation feature pyramid network is introduced to fuse shallow and deep semantic information to improve the detection effect of the network on small targets. Then, GHMC and GHMR loss functions are used to classify and locate defects. Finally, the soft-non maximum suppression algorithm in Gaussian form is introduced to improve the detection accuracy. The experimental results show that the average accuracy of the improved RetinaNet-GHM algorithm is 76.7%, and the average accuracy of crazing, inclusion, patches, pitted surface, rolled-in_scale and scratchs is 45.2%, 88.2%, 94.2%, 86.1%, 65.1% and 87.4% respectively. Compared with other classical algorithms, the improved RetinaNet-GHM algorithm has better detection effect.