In China, citrus production occupies an important position in agriculture and has great economic benefit. For a long time, most of citrus harvesting relies on manual work, which has low efficiency and high labor cost. The labor cost accounts for almost onehalf of total labor cost in citrus production process. In addition, citrus picking is usually carried out during the day, while makes less use of night time. Therefore, it is of great significance to develop a fruit picking robot working at nighttime. Focusing on citrus picking process, a multiscale convolution neural network named Des-YOLO v3 was proposed and used to detect citrus at nighttime under natural environment. By using ResNet and DenseNet for reference, the Des-YOLO v3 network was designed to realize the reuse and fusion of multilayer features of the network, which strengthened the robustness of small target and overlapping occlusion fruit recognition, and significantly improved the precision of fruit detection. The experimental results showed that the precision, recall rate and F1 value of the Des-YOLO v3 network were 97.67%, 97.46% and 0.976, respectively, while those of YOLO v3 network were 91.41%, 91.10% and 0.913, respectively. At the same time, the mean average precision of the trained model under the test set was 90.75%, and the detection speed was 53f/s, which was 2.27 percentage points and 11f/s higher than those of YOLO v3_DarkNet53, respectively. The final results showed that the Des-YOLO v3 recognition network had stronger robustness and higher detection precision for the recognition of mature citrus in the complex field environment at night, which provided technical support for the visual recognition of citrus picking robot.