A fine-grained convolution module SPD-Conv was proposed to replace the convolution subsampling for YOLOv8s network and extract the features of small defects in a fine-grained way in order to improve the accuracy and recall rate of the detection of small defects on the surface of drum rollers and enhance the detection ability of the model for small target defects. GFPN feature fusion module was introduced to enhance the cross-scale connection between adjacent layers and cross-layer connection under the same scale in the feature fusion module, which is conducive to the transmission of small target feature information in the convolutional network. The small target detection layer was added to the head in order to improve the detection ability of the model. The boundary frame loss function of Wise-IOU was used to replace CIOU in terms of loss function, which could accelerate network convergence and improve the accuracy of network detection. The test was conducted on the self-made drum roller defect dataset. Results showed that the improved YOLOv8s achieved 0.911, 0.983 and 0.935 in the chamfer dataset, side dataset and end dataset, respectively. mAP@0.5 increased by 6.4%, 3.3% and 4% respectively compared with YOLOv8s. Accuracy and recall rates have improved with an average detection time of 23 ms per image. The improved YOLOv8s has better localization ability and detection accuracy for small target defects compared with the original model, and the detection speed can meet the requirements of industrial mass detection.