Abstract: The current potato grading methods mainly rely on manual grading, which suffers from low efficiency, high subjectivity, high cost, and insufficient detection accuracy. These limitations make it difficult to meet the demands of large-scale production and high-precision detection. Additionally, potato surface defect detection faces several challenges. First, the variety of defects, such as disease spots, pest damage, and mechanical injuries, along with their diverse morphologies, makes it difficult to achieve effective recognition using simple image processing methods. Second, the complex environmental factors in agricultural settings, such as changes in lighting and soil coverage, further increase the detection difficulty. Moreover, most existing deep learning models have large parameter sizes and high computational complexity, making them unsuitable for real-time operation on resource-constrained devices.To address these issues, this paper proposes a potato surface defect detection method, DATW-YOLOv8, based on an improved YOLOv8n model. The algorithm replaces the Bottleneck module in C2f with the Dilation-wise Residual (DWR) module and introduces the Dilated Reparam Block (DRB) to optimize the extraction of detailed features, thereby improving the accuracy of defect feature detection. Additionally, a lightweight adaptive downsampling (ADOWN) convolution module is integrated to achieve efficient dimensionality reduction and enhance the model's processing efficiency. The detection head is upgraded to a Task Align Dynamic Detection Head (TADDH) to improve the accuracy of defect boundary prediction and focus precisely on key defect regions. Finally, Wise-EIoU is adopted as the bounding box regression loss function to increase the model’s attention to difficult samples, thereby enhancing boundary regression accuracy and model robustness.Experimental results show that the improved DATW-YOLOv8 model achieves detection accuracy, recall, and mean average precision (mAP) of 95.8%, 88.1%, and 94.3%, respectively. The parameter size and weight size are 1.5M and 3.6MB, which are 50.0% and 42.9% smaller than those of the original YOLOv8n model. Additionally, the accuracy, recall, and mAP are improved by 2.8%, 1.6%, and 1.4%, respectively. When using different downsampling methods for defect detection, ADOWN exhibits significant advantages in terms of detection accuracy and mAP compared to YOLOv7 E-ELAN, SPDConv, WaveletPool, and Light-weight Context Guided DownSample, with mAP improvements of 0.8%, 1.3%, 1.4%, and 1.7%, respectively, while reducing parameter counts by 11.8%, 60.5%, 6.3%, and 48.3%, respectively. Additionally, Wise-EIoU achieves the lowest loss value, the fastest convergence, and the smallest fluctuation among various bounding box loss functions.In comparison with YOLOv5Lite-g, YOLOv8n, YOLOv10n, and Mobilenetv2-SSD, DATW-YOLOv8 outperforms these models in terms of accuracy, recall, and mAP@0.5, while maintaining a lower model weight than other network models. Compared with YOLOv7tiny, the model weight of DATW-YOLOv8 is only one-third of that of YOLOv7tiny, further demonstrating its superior performance across multiple evaluation metrics. Furthermore, the improved DATW-YOLOv8 model was deployed on a potato surface defect online detection and sorting test bench to conduct online detection experiments on different potato varieties and conveyor speeds. The results indicate that the maximum sorting accuracy reaches 95.8%, effectively meeting the real-time detection requirements for potato surface defects and demonstrating its potential for practical production applications. Overall, this method provides a valuable technical reference for online potato surface defect detection and model deployment on mobile devices.