Abstract: In the field of deep learning, the generalization ability of models is often significantly compromised when the original data of tea leaves is insufficient, leading to a substantial decline in the detection capability for tender tea shoots. To address this issue, this study proposes a Tea DCGAN (tea deep convolution generative adversarial networks) and its corresponding data augmentation method. Initially, a 64×64×64 layer was added to both the generator and discriminator of the DCGAN (deep convolution generative adversarial networks) to enhance the model's perception and learning ability for low-dimensional features. Additionally, the LeakyReLU (leaky rectified linear unit) function in the DCGAN was replaced with the more linearly controllable ELU (exponential linear units) function, thereby improving the stability and accuracy of model training. Subsequently, a Fake Tea data augmentation algorithm framework was constructed based on the Tea DCGAN network. This framework analyzes the distribution of real tender tea shoots in the existing dataset to understand the underlying patterns. According to these patterns, the sample images generated by the Tea DCGAN network are distributed into real tea tree images, automatically forming a deep learning dataset. Finally, the proposed data augmentation method underwent several experiments, including adversarial generative network ablation tests, rare tea variety control tests, and comparisons of various data augmentation methods at different scales. The ablation test results indicated that Tea DCGAN performed optimally in terms of the FID (Frechet Inception Distance) metric, especially after 100,000 training epochs, where the FID value for the Zijuan tea variety dropped from 322.10 to 265.63, and for the Longjing 43 tea variety, it decreased from 396.38 to 323.09, significantly enhancing the quality of the generated images.In various detection model experiments with multiple data augmentation methods, the Fake Tea method outperformed other approaches across different detection models. Specifically, the Faster R-CNN model achieved an mAP of 42.71% and 38.46% on datasets comprising 25 Longjing 43 and 25 Zijuan tea varieties, respectively. As the dataset size increased, the performance of all methods improved, but the Fake Tea method consistently maintained the highest mAP value across all dataset sizes. Notably, when the original dataset consisted of 200 images, the mAP value reached 89.41%, making it suitable for intelligent tea harvesting.The findings of this study demonstrate the effectiveness and superiority of Tea DCGAN and the Fake Tea data augmentation method in tea leaf image generation and object detection tasks. The proposed Tea DCGAN and Fake Tea data augmentation methods can effectively alleviate difficulties in data acquisition and the scarcity of samples, significantly enhancing the accuracy of tender tea shoot detection in scenarios with limited samples.