Abstract: Abstract: The surface defect detection of titanium alloy connectors is very important for the safety and reliability of aviation equipment. However, the existing deep learning models have problems such as difficulty in taking into account multi-scale defects, confusion between defects and background, and easy loss of fine-grained features in the face of connectors with complex morphology. In order to solve the above problems, this paper proposes a new model architecture DEIM-MFE based on the DEIM-RTdetr model. Firstly, the real images of titanium alloy connectors in the actual industrial scene are collected, and a special data set (TAC-Dataset) for surface defects of titanium alloy connectors for aero-engines is constructed. Secondly, a new encoder structure, multi-scale feature enhanced encoder (MFE Encoder), is proposed. The multi-scale feature alignment and cross-attention fusion module (MFAC) enhances the fusion ability of multi-scale semantic information by performing spatial alignment and cross-attention interaction on different levels of features. The local residual enhancement module (LRE) uses convolution operation to strengthen local texture and edge information, and improves the model 's ability to distinguish between defect areas and complex backgrounds. The shallow feature recovery module (SFR) is guided by the output multi-scale enhanced semantic features, and deeply reconstructs the shallow features through a multi-branch aggregation structure, so that the details and semantics fully interact, thereby enhancing the ability to characterize small defects. The experimental results on TAC-Dataset show that compared with the benchmark model, the proposed model improves Precision, Recall, F1-Score and mAP @ 0.5 by 2.37, 1.73, 2.10 and 2.02 percentage points respectively. This model shows good recognition and positioning ability in the surface defect detection task of complex connectors, and provides a reliable and effective technical solution for the surface quality detection of key connectors.

Key words: titanium alloy connectors, surface defects detection, DEIM, multi-scale features, Cross-Attention

摘要： 摘 要: 钛合金连接件表面缺陷检测对于航空装备安全与可靠性至关重要，然而现有的深度学习模型在面对具有复杂形貌的连接件时存在多尺度缺陷难以兼顾、缺陷与背景易混淆、细粒度特征易丢失等问题。为解决上述问题，本文在DEIM-RTdetr模型的基础上提出了一种新的模型架构DEIM-MFE。首先，采集实际工业场景中真实的钛合金连接件图像，构建面向航空发动机的钛合金连接件表面缺陷专用数据集（TAC-Dataset）。其次，提出一种新的编码器结构-多尺度特征增强编码器（MFE Encoder），其中，多尺度特征对齐与交叉注意力融合模块（MFAC）通过对不同层级特征进行空间对齐与交叉注意力交互，增强多尺度语义信息的融合能力；局部残差增强模块（LRE）利用卷积操作强化局部纹理与边缘信息，提高模型对缺陷区域与复杂背景的区分能力；浅层特征恢复模块（SFR）以输出的多尺度增强语义特征为引导，通过多分支聚合结构对浅层特征进行深度重构，使细节与语义充分交互，从而增强微小缺陷表征能力。在TAC-Dataset上实验结果表明，相较于基准模型，所提模型在Precision、Recall、F1-Score、mAP@0.5上分别提升了2.37、1.73、2.10、2.02个百分点。本模型在复杂连接件表面缺陷检测任务中表现出良好的识别与定位能力，为关键连接件表面质量检测提供了可靠有效的技术方案。

关键词: 钛合金连接件, 表面缺陷检测, DEIM, 多尺度特征, 交叉注意力