As a kind of special equipment, the safety of pressure vessels attracts more and more attention. To ensure their safe operation, using the pressure vessel crack image as the research object, this paper constructed an algorithm model for the crack detection and identification. Generally, the algorithm model is constrained by various hardware conditions, such as memory space and processor computing power during actual deployment. Therefore, a lightweight method based on NewEfficientNet-B0 was proposed. The results show that the algorithm model reduces the number of model parameters by 78%. To deal with the difficulty of recognizing tiny crack images, an improved multi-scale prediction method was proposed. The detection and recognition accuracy rate of 81% was achieved on the test data set.