Aiming at the problems of rising labor costs and an aging population, it is crucial to explore mango recognition technology and edge computing model deployment methods based on lightweight deep learning models to promote the development of lightweight and low-cost mango picking robots. A mango recognition method based on the improved YOLO 13n model was proposed by using YOLO 13n as the baseline model, and it was deployed on edge computing devices. Firstly, the SE module was introduced into the backbone network of YOLO 13n to improve the feature expression ability of the model. Secondly, a CBAM module was added to the neck network to fuse more feature information and strengthen the channel features of mango recognition, and highlight the mango region in the image. Then, the introduced SE module and CBAM module were used to reshape the YOLO 13n architecture to obtain the improved YOLO 13n model, and finally, the deployment of the improved YOLO 13n model on edge computing devices was realized by comparing the model deployment methods. The experimental results showed that the mango detection performance of the improved YOLO 13n model was P was 94.5%, R was 91.2%, AP0.5 was 95.6% and AP0.75 was 94.9%, which exceeded many lightweight models, including YOLO v8n, YOLO v9s, YOLO v10n, YOLO 11n, YOLO 12n and YOLO 13n. In addition, compared with the original YOLO 13n model, the P, R, AP0.5 and AP0.75 of the improved YOLO 13n model were improved by 0.6, 0.6, 1.1 and 1.5 percentage points, respectively, and the lightweight characteristics were maintained. To verify the effectiveness of the improved YOLO 13n model on edge computing devices, the improved model was deployed on the NVIDIA Jetson Orin Nano. The maximum inference speed for a single image reached 31.71 f/s. Compared with the original YOLO 13n model, the inference speed only decreased by about 0.4 f/s, still exceeding 30 f/s, meeting the real-time demand for mango recognition. By measuring the inference time on 50 randomly selected sample images, the average inference time was 36.88 ms, indicating the stability and reliability in real-time mango recognition. The research result can provide technical support for the development of lightweight, low-cost mango picking robots.