ABSTRACT Real‐time and efficient surface roughness detection is crucial in modern manufacturing. Noncontact measurement based on machine vision has attracted considerable attention owing to its high efficiency and strong adaptability. However, the reflection characteristics of the same surface vary under different light sources. To address this issue, a roughness dataset is constructed using zirconia, alumina, and ZTA materials, which is imaged under various light sources (white, red, green, and blue). We propose a novel bimodal feature fusion method based on the YOLOv8 network, termed DAM‐YOLO. Specifically, we construct a bimodal feature fusion module (DAM) that integrates the channel attention mechanism (CAM) and the spatial attention mechanism (SAM). Leveraging bimodal image inputs, DAM enhances the feature maps of single modalities by exploiting the intermodal dependency relationships, thereby strengthening the mixed spatial‐channel weights of cross‐modal fusion outputs. By fusing cross‐modal complementary features to refine the original unimodal features, DAM enables the formation of more discriminative features, ultimately generating highly representative bimodal feature maps. Furthermore, we integrate the online convolutional re‐parameterization (OREPA) and generalized efficient layer aggregation network (GELAN) modules into the detection network. This integration ensures that the model's feature extraction capability is significantly enhanced under the condition of low parameter complexity. Extensive experiments validate that the proposed DAM‐YOLO method achieves a mean average precision (mAP) of 0.993, which fully demonstrates the accuracy and effectiveness of our approach.
Du et al. (Fri,) studied this question.