Abstract: Pressurized water reactor nuclear fuel rods are a type of fuel element commonly used in nuclear power plants. The production quality of fuel rods is related to the safe operation of nuclear power plants. It is very important to inspect the appearance quality of fuel rods. In view of the complex background and difficult feature extraction of fuel rod appearance defects in the produce scene, this paper proposed a novel cross stage fusion model based on dual dimensional self attention (DSCFM) . The model used Extended-ELAN as the backbone network to extract features and designed a novel multi-scale feature fusion structure (MFFS) as the neck structure of the model to efficiently process and fuse feature information at different levels. The purpose of the MFFS design is to optimize the information flow between features at different levels. A large amount of detailed information was retained through a detailed fusion mechanism, while strengthening the model’s ability to understand the complexity of the scene. In addition, to further mine and utilize the underlying structural information and effectively integrate deep information, a self attention feature fusion module with dual dimensional characteristics (DSAF) was proposed. DSAF expanded the deep and underlying features into dual dimensional feature maps, and used its own transposed matrix to generate channel and spatial attention maps, thereby accurately enhancing the expression of key information, while suppressing irrelevant noise and optimizing the feature fusion process. Through this dual dimensional self attention mechanism, DSAF dynamically adjusts feature responses, effectively captures long-term dependencies, and enhances the model’s adaptability and interpretation capabilities for complex scenes. Finally, combined with a multi-scale deconvolution structure, the DSCFM achieves effective upsampling and optimization of features, significantly improving the model’s ability to capture information at different scales and its robustness in various visual tasks. The results are verified on a fuel rod appearance defect dataset, and the experiments show that compared with other detection models, the DSCFM can quickly and accurately identify defects, with an mAP of 82.0% and a recall rate of 77.9%.