【摘要】 汽车发动机铸件的表面缺陷给汽车出行带来了重大安全隐患。由于汽车发动机铸件结构复杂、缺陷种类多、部分缺陷面积较小，且多种缺陷形态特征与铸件结构特征相似，使用现有检测方法对其识别时检测效率低，误检、漏检问题频发。为解决上述问题，提出一种基于优化残差和融合注意力的铸件缺陷检测网络。首先，网络通过构建的优化残差模块，保证表面缺陷纹理特征完整性并有效降低了模型的参数量和计算量；其次，构建双层注意力融合机制，提高了对铸件表面缺陷特征的表征能力，降低对无关背景信息的关注度；最后，结合跨尺度加权特征融合模块从多尺度、多维度展开目标预测，同时，针对CIOU LOSS存在的不足加以改进，加强损失函数对表面缺陷边界框的精准定位。实验结果表明所提方法在各项评价指标上均取得良好效果，全类平均检测精度达92.28%,单图推理时间达到了13.8 ms,实现了汽车发动机铸件表面缺陷的高效、精准检测。更多还原

【Abstract】 The surface defects of automobile engine castings bring great safety risks to automobile travel. Due to the complex structure of automotive engine castings, the variety of defects, the small area of some defects, and the shape characteristics of many defects are similar to the structural characteristics of castings, the detection efficiency is low when the existing detection methods are used to identify them, and the problems of false detection and missing detection are frequent. To solve the above problems, a casting defect detection network based on optimization residuals and fusion attention was proposed. Firstly, the network ensures the integrity of texture features of surface defects and effectively reduces the number of parameters and calculation amount of the model by constructing an optimized residual module. Secondly, the two-layer attention fusion mechanism is constructed to improve the characterization ability of casting surface defects and reduce the attention to irrelevant background information. Finally, combined with the cross-scale weighted feature fusion module, the target prediction was expanded from multi-scale and multi-dimension. At the same time, the shortcomings of CIOU LOSS were improved to strengthen the accurate location of the surface defect boundary box by the loss function. The experimental results show that the proposed method achieves good results in all evaluation indexes, the average detection accuracy of all classes is up to 92.28%, and the reasoning time of single graph is up to 13.8 ms, which realizes the efficient and accurate detection of automotive engine castings surface defects.更多还原