Addressing the issues of high impurity content, low cleaning efficiency, and high labor intensity in the collection and storage of potatoes, the use of intelligent cleaning robotic arms to replace traditional cleaning tasks has become an inevitable trend. The vision component of these systems is crucial for the precise real-time positioning and high-quality identification and detection of potatoes. This study introduces a YOLOv8-based method for detecting and categorizing stored potatoes, named YOLOv8-PD. Firstly, ODConv is introduced into the C2f module to enhance the recognition accuracy of complex potato shapes and textures through a multi-dimensional attention mechanism, while suppressing background interference. Secondly, A weighted BiFPN is integrated to enhance multi-scale feature fusion. Finally, the CARAFE upsampling operator replaces YOLOv8’s UpSample. Improving detail preservation and reconstruction quality during upsampling and selecting important feature information. Experimental results show that, on our self-constructed Potato Detection dataset, The enhanced algorithm shows improvements of 2.2%, 1.4%, and 0.4% in precision (P), recall (R), and mAP, respectively, compared to the original. The model’s representation capability is enhanced, and the algorithm’s detection speed reaches 145.16 FPS, enabling real-time and accurate differentiation between potatoes and other categories. The proposed algorithm improves the accuracy of potato recognition and positioning, providing significant reference value for the study of automated picking and cleaning processes for stored potatoes.