With the continuous development of manufacturing technology and structural design, additive manufacturing technology and metal lattice structure occupy a strategic position in the entire aerospace industry chain. Inside the metal three-dimensional lattice structure manufactured by additive manufacturing technology, there may be some defects such as cracks, incomplete fusion, chasm, etc. The defects will reduce the structure-function performance of the metal lattice structure. So it is necessary to carry out corresponding non-destructive testing technology research. Based on the gray value distribution characteristics of computed tomography (CT) scanning images, a method for detecting the internal defects of metal three-dimensional multi-layer lattice structures is proposed. Combining industrial CT scanning with convolutional neural network, a feature learning network based on the Faster R-CNN network architecture is designed. The defects in the obtained grayscale image are detected and identified to determine their specific positions. The experimental verification results show that, compared with the manual marking method, the recognition rate of typical internal fault defects of metal three-dimensional multi-layer lattice structure samples reaches 99.5%.