Under the same detection frequency and depth, when the center spacing of multiple defects is less than the resolution threshold determined by the Rayleigh criterion, it is challenging to achieve super-resolution imaging of multiple defects using the ultrasonic Total Focusing Method (TFM). A multilevel deep learning network is proposed as a super-resolution reconstruction method for ultrasonic Lamb wave TFM images . The first-level network is a detection network that uses a Resnet50 with more convolutional layers to improve the linear expression capabilities of the neural network. It introduces residual structure to solve low accuracy issues of multi-convolutional layers so that the Resnet50 can accurately detect defects from TFM images. The second-level network is a super-resolution reconstruction network that uses a Deeplab v3+ with a dilated convolutional layer. It controls the receptive field without changing the feature map size of the TFM image. With this model, a super-resolution reconstruction of multiple defects with a center spacing less than the resolution threshold is realized by extracting the detailed features of defects in the TFM image. Experimental results show that when the defect center spacing is greater than the resolution threshold determined by the Rayleigh criterion, the super-resolution reconstruction method improves the calculation accuracy of the defect center spacing by 4.7% and the calculation accuracy of the defect area by 93.7% compared with TFM. When the defect spacing is less than the resolution threshold, the method can still identify and accurately calculate the center spacing of multiple defects.