To enhance the picking efficiency and accuracy of tea picking robots while ensuring compatibility with mobile devices with limited computational power, this study proposes a tea bud recognition method based on an improved YOLOv7—tiny model. The HardSwish activation function was incorporated to minimize accuracy loss of numerical values in quantization mode and improve the model's precision. Furthermore, the ELAN module in the original network was optimized using depthwise separable convolution, effectively reducing parameter redundancy and computational requirements in the module, thereby making the model more lightweight. To mitigate the accuracy loss associated with model lightweighting and to further improve the model accuracy, the EIoU bounding box loss function, known for its sensitivity to small objects, was employed. Experimental results showed that the improved model achieved P, R, and mAP of 79.6%, 79.1% and 81.4%, respectively, which represented improvements of 5.4%, 2.3% and 2.7% over the original model. Additionally,the parameter quantity of the improved model was reduced to 4.8 M, which was a 20.0% decrease compared tothe original model, while computational complexity dropped to 10.8 GFLOPs, marking a 16.9% reduction.