Abstract: Residual cores in aero-engine blades pose significant safety risks, as their presence compromises structural integrity by obstructing cooling channels, degrading material properties (e.g., creep resistance), and impairing film cooling efficiency, thereby escalating the potential for thermal distress and catastrophic failures. Given their non-destructive nature and sensitivity to light elements, neutron imaging has emerged as a viable technique for detecting such defects. However, traditional reliance on manual interpretation of neutron images is plagued by inherent limitations, including human subjectivity, inconsistent accuracy, and labor-intensive processes. To address these deficiencies, an enhanced automated detection framework was proposed, integrating an optimized YOLOv11 deep learning model with data augmentation strategies. The primary contributions and methodologies are as follows: First, to mitigate the challenge of low contrast inherent in neutron images, contrast limited adaptive histogram equalization (CLAHE) was systematically applied. This technique adaptively normalizes pixel intensities within localized image tiles while constraining contrast amplification, effectively enhancing defect visibility while preserving structural details. The CLAHE-preprocessed images serve as the foundation for training the deep learning model, ensuring robust feature extraction. Second, to improve the model’s sensitivity to small-scale residual core defects (≤0.5 mm), the YOLOv11 architecture was refined through two key modifications: integration of an asymptotic feature pyramid network to hierarchically fuse multi-scale feature representations for enhanced contextual awareness, and implementation of a spatial and channel synergistic attention module to dynamically weight critical spatial regions and feature channels, thereby suppressing irrelevant background noise. This hybrid architecture enables precise localization even in complex blade geometries. Finally, to overcome data scarcity (a pervasive challenge in neutron imaging applications), a novel data augmentation strategy was devised. Synthetic training samples were generated by adhering gadolinium-doped ceramic fragments to pristine blades, leveraging gadolinium’s high neutron attenuation properties to simulate realistic defect signatures. The resulting dataset encompasses diverse defect morphologies, sizes, orientations, and spatial distributions, ensuring model generalizability across heterogeneous real-world scenarios. Experimental evaluations, conducted on a curated dataset of neutron images, demonstrate the superiority of the proposed framework. The enhanced model achieves a precision of 95.0%, recall of 91.4%, mAP@0.5 of 95.8%, and mAP@0.5:0.95 of 79.3%, outperforming baseline methods. These metrics collectively validate the model’s efficacy in detecting minute defects while maintaining high reliability. The study’s implications extend beyond aero-engine inspection, offering a scalable solution for automated defect detection in neutron imaging applications across safety-critical industries. By automating this quality control bottleneck, the framework holds promise to accelerate manufacturing efficiency, eliminate human variability, and facilitate real-time monitoring, thereby advancing the adoption of Industry 4.0 principles in aerospace and related domains. The synergistic integration of advanced image preprocessing, attention-driven deep learning architectures, and physics-informed data augmentation strategies establishes a new paradigm for non-destructive evaluation of composite metallic components.