Tire internal defect detection can effectively identify potential issues during the manufacturing process, providing strong support for process adjustments and ensuring driving safety. Defect targets in tire X-ray images are characterized by multi-scale features, extreme aspect ratios, diverse and irregular shapes, a large number of small targets, and an imbalance between positive and negative samples, which results in low detection accuracy. To address these challenges, we propose a tire defect detection method based on an efficient encoder and multi-scale feature fusion. First, an efficient encoder is designed by combining deformable attention mechanisms and channel attention mechanisms to enhance feature extraction and representation capabilities. Then, a multi-scale feature extraction and fusion module is constructed to integrate shallow and deep feature information, preserving critical contextual information and improving feature representation diversity. Finally, an adaptive bounding box regression method is employed during model training to dynamically allocate weights to samples based on difficulty, reducing the impact of invalid samples and achieving faster model convergence while enhancing generalization. Experimental results demonstrate that the proposed model achieves a mean average precision (mAP) of 95.5% on the tire defect dataset, a 3.6 percentage point improvement over the baseline network, thus laying a solid foundation for the practical application of tire defect detection.