To detect the possible defects of metal castings in engineering and application,the laser ultrasonic visual detector based on thermoelastic mechanism is used to scan the castings and make a maximum amplitude image of the signal to realize the visual detection of the castings.In order to efficiently and quickly process the maximum amplitude graph in batches,the convolution neural network image processing technology is used to identify the maximum amplitude graph.A convolutional neural network architecture is designed to recognize the maximum amplitude figure,and different convolutional neural network architectures are set up by changing the size of convolutional layers and convolutional kernels during the recognition process,and the predefined architectures are horizontally compared with other architectures.The experimental results show that the overall performance of predefined architecture is the best.Under the same experimental conditions,the convolutional neural network architecture provides an effective and feasible scheme for casting defect detection using maximum amplitude graph.