After the potato seed tuber cutting machine completes the cutting process, it is difficult to ensure that all the potato tuber pieces contain bud eyes. Therefore, it is necessary to conduct bud eye detection on the potato tuber pieces to achieve automatic screening of potato tuber pieces without bud eyes. Due to the different cut-ting methods, the potato tubers cut have differences in morphology, which poses higher requirements for the generalization ability of the potato tuber bud eye detection model. To address these issues, this study proposes a potato tuber seed tuber bud eye detection model based on the improvement of YOLOv8s, named YOLOv8s-LEW. Firstly, the lightweight shared convolution detection head LSCD is integrated to enhance the infor-mation interaction across hierarchical levels, and a more efficient feature transmission path is constructed to optimize the overall detection performance of the model; Secondly, the C2f_EVA embedding model backbone network is designed based on the idea of the efficient visual attention module EVA, and its large-core atten-tion mechanism can expand the effective receptive field of the network, enabling features to cover a larger spatial context information, thereby enhancing the model's ability to discriminate the details of bud eyes. Fi-nally, WIoU-v3 with the non-monotonic focusing mechanism is selected as the bounding box loss function, and the training weights of different quality samples are dynamically regulated to improve training efficiency and detection accuracy. The model is trained on the same potato tuber bud eye dataset, and the experimental results show that the improved model YOLOv8s-LEW has an accuracy, recall rate, and average precision that are 2.32%, 2.42%, and 2.61% higher than the original model, respectively, reaching 97.93%, 90.49%, and 95.86%, and the model parameters and floating-point operation quantities are reduced by 10.6% and 6.0%, respectively. This indicates that the model improves the bud eye detection performance while maintaining a certain level of computational efficiency, has better comprehensive detection performance, and can provide technical reference for the automatic screening of potato tuber seed tubers.