Titanium alloy additive manufacturing is one of the disruptive technologies in the manufacturing industry, and titanium alloy products serve as critical materials for national defense equipment such as aircraft and missiles. During production and service, titanium alloy products are susceptible to surface defects such as cracks and scratches caused by harsh external environments. If such defects are not detected and addressed promptly, they may lead to serious consequences. Therefore, surface defect detection for titanium alloys is an important task. However, current detection methods suffer from low efficiency and several other limitations, including difficulties in deployment on terminal devices. To overcome these shortcomings, this paper proposes Titanium-YOLO, an improved surface defect detection algorithm for titanium alloys based on the YOLOv10n model.

To improve the YOLOv10n base model, key literature was reviewed to identify its fundamental network structure, which comprises the backbone, neck, and detection head. It was determined that developing an efficient detection algorithm relied on multifaceted lightweight improvements to this base structure. Guided by convolutional neural network principles, DualConv was introduced to replace certain standard convolutions in the backbone, and the traditional C2f module was replaced by the C2f-DualConv module. In the neck, the SlimNeck structure was adopted for lightweight enhancement, incorporating key lightweight modules such as GSConv and VoVGSCSP. For the detection head, the lightweight litehead replaced the traditional v10Detect, where DWConv and SPConv substituted standard convolutions to reduce module redundancy. Finally, using a self-constructed titanium alloy dataset and existing laboratory experimental conditions, ablation experiments were conducted to evaluate the performance of each improved module, alongside comparative experiments evaluating the Titanium-YOLO algorithm with mainstream YOLO algorithms.

Results of the ablation experiments (Table 4, Fig.13) and comparative experiments (Table 5) demonstrated that the improved Titanium-YOLO model had 2.6×106 parameters, a reduction of 3.7% compared to the baseline model, and a computational cost of 7.3×109, which was 13.1% lower. Meanwhile, the improved algorithm achieved a mean average precision (mAP) of 99.0% on the self-constructed titanium alloy dataset, surpassing the baseline by 2.4 percentage points. Furthermore, compared with mainstream YOLO algorithms such as YOLOv5, YOLOv7, and YOLOv8, the improved Titanium-YOLO algorithm demonstrated higher accuracy and a more lightweight design.

Timely detection of surface defects in titanium alloys is crucial to ensure component safety and prevent serious incidents. Nonetheless, current detection technologies suffer from low efficiency and challenges in deployment on terminal devices. This paper proposes a lightweight Titanium-YOLO detection algorithm, improved from YOLOv10n through the comprehensive integration of DualConv, the SlimNeck structure, and the litehead module. Experimental results show that the Titanium-YOLO algorithm achieves a mAP of 99.0% on the self-constructed titanium alloy dataset, with the number of parameters and computational cost reduced to 2.6×10⁶ and 7.3×10⁹, respectively. The improved algorithm significantly outperforms the baseline and other mainstream models, offering higher accuracy, smaller model sizes, lower computational cost, and simpler deployment. Future work will focus on continued iteration and optimization of the algorithm, with practical implementation and deployment of detection equipment becoming a key research focus.