In order to meet the rapid and accurate recognition requirements of citrus picking robot, a citrus recognition method based on improved YOLOv3-tiny（You Only Look Once v3 tiny） light convolutional neural network model was proposed. The proposed method replaces the original loss function of YOLOv3-tiny convolutional neural network model with DIOU（Distance Intersection over Union） loss function to improve the identification accuracy. Mobilenetv3-small convolutional neural network was used to replace the backbone feature extraction network, which made the model lighter and improved the recognition speed. A new residual structure is added in Mobilenetv3-small to reduce the loss of characteristic information and improve the identification accuracy. Simplified SPP（Spatial Pyramid Pooling） network structure and separable convolution layers are added to the backbone feature feature extraction network to improve the feature information extraction capability of the model. A down-sampling layer is used to integrate feature information of two scales, and the Hard Swish activation function is also applied. Proposed method was compared with the recognition effect of YOLOv3-tiny on citrus data set, the average mAP and F1 values of proposed method were 96.52% and 0.92, increased by 3.24% and 0.03, respectively. The recognition time cost and model weight were only 47 ms and 16.9 MB, respectively, which reduced by 24% and 49%. Further experiment was conducted to verify the recognition accuracy of citrus under different environmental conditions, The accuracy of YOLOv3-tiny were 98.6%, 90.5%, 95.8% and 86.8% respectively under sufficient light and no shade, insufficient light and no shade, and increased by 0.7%, 6.5%, 3.2% and 7.7%, respectively. The results showed that the improved YOLOv3-tiny light citrus recognition method achieves high recognition accuracy, fast recognition speed and light weight.