Abstract: Neutron detection is an important technology in the field of nuclear energy development and is involved in many research and application areas, such as particle physics, material science, cosmic ray detection and even environmental monitoring, oil well detection and nuclear medicine, etc. Since neutron scintillator detectors often respond to both neutron and γ-ray, effective discrimination between neutron and γ-ray is a prerequisite for high-precision neutron detection. In order to further explore way to enhance the performance of n/γ discrimination, this paper combined the pulse shape discrimination (PSD) technique and the Gramian angular field (GAF) image transformation method, and applied the convolutional neural network (CNN) classification model to the n/γ discrimination work. The ²³⁹Pu-Be neutron source and the Cs₂LiYCl₆:Ce³⁺ (CLYC) detector were used to set up an experimental platform for the n/γ hybrid radiation field, and 20 000 original pulsed one-dimensional sequence samples were acquired through the Tektronix model DPO4034 oscilloscope. In the experiment, the charge comparison method was adopted to discriminate the original samples, and the discrimination results can be used to produce the labels of the dataset used in the GAF-CNN method and for the final comparison of the discriminative performance of the various methods. Due to the excellent performance of the CLYC detector, the discrimination effect of the charge comparison method is good, which ensures that the labeling of the dataset can be produced with high accuracy, and after the best performance of the charge comparison method has been achieved through the optimization of the window, the gap between the upper limit of the performance of the traditional method and the neural network method can be clearly found. The GAF-CNN discrimination method transformed the n/γ pulse data into a two-dimensional image through the GAF, after which the image was fed into the CNN classification model for sample discrimination, which transformed the n/γ discrimination problem into a simple image binary classification problem. Since the nuclear pulse signal is a typical time series, the use of GAF can retain the time domain features of the nuclear pulse in a more complete way, and the convolution operation of the CNN can utilize the frequency domain features, so the GAF-CNN is a kind of discrimination method that can utilize the time-frequency features at the same time. In order to verify the accuracy of GAF-CNN discrimination, the discrimination effect was compared with the traditional CNN discrimination method and the charge comparison method, where the traditional CNN discrimination method refers to the method of simply collapsing the pulsed one-dimensional sample sequences into a two-dimensional matrix and inputting it to the CNN for sample identification. The results show that the GAF-CNN discrimination method has a lower discrimination error rate and shorter processing time, and the figure of merit (FOM) of n/γ discrimination has an order of magnitude improvement. Meanwhile, it has the characteristics of network lightweight, which helps to realize the embedded deployment of convolutional neural network PSD algorithm, and provides a feasible PSD technology solution for the development of high-performance n/γ composite detection spectrometer.