Rubber bushings, as a flexible connection widely used in automotive suspension systems to meet the requirements of noise reduction and vibration damping, are prone to various surface defects during the production process. Regarding the detection of two typical types of defects, there exist issues such as high rates of missed targets and difficulties in meeting real-time online inspection requirements of enterprises. In response to the aforementioned issues, a small object detection model based on YOLOX is proposed. Firstly, a hardware testing platform was established to study two typical surface defect characteristics of rubber bushings and create a dataset. Secondly, Res2Net_MSDA module was introduced in the backbone network to bolster its capability in extracting features across multiple scales more precisely. Subsequently, the feature fusion layer was augmented with ACmix module, broadening the model's capacity for comprehensive understanding and strengthening its proficiency in capturing essential data from key areas. Lastly, an optimal hyperparameter search strategy based on the Particle Swarm Optimization (PSO) algorithm was constructed. The experimental findings reveal that the YOLOX-PSO model attains a mean accuracy of 87.26% on the test dataset, with F1 scores of 0.99 and 0.76, and a Frames Per Second (FPS) value of 42.5. Compared to mainstream deep learning models, the YOLOX-PSO model exhibits improved detection accuracy while maintaining considerable efficiency, demonstrating its effectiveness and reliability.