【目的】针对已有模型在检测小尺寸、与背景对比度低的破损时易出现漏检、误检等问题,提出一种改进YOLOv4的检测模型。【方法】针对小尺寸问题,首先设计了DDSunit替换主干网络中的Resunit,利用不同层次特征跨层连接的方式获得完整丰富的多尺度特征完成小尺寸破损的检测。其次,在分类损失函数中引入梯度协调机制动态调整小尺寸破损的权重使其得到充分训练。针对破损与背景对比度低的问题,首先在主干网络深层网络层中嵌入坐标注意力机制,增强模型对破损特征的关注度,降低背景噪声的干扰。其次,设计精确解耦头通过解决分类、定位任务对特征需求的矛盾提升检测精度。【结果】实验表明,模型的平均精度较YOLOv4提升了3.92%,小尺寸的划伤类破损、与背景低对比度的磨损类破损的检测精度分别提升4.32%和4.24%,可有效解决漏检、误检问题。

【Purposes】An improved YOLOv4 detection model is proposed to solve the problems  of missing detection and false detection when the existing models detect objects with small size and  low contrast with the background. 【Methods】In order to solve the problem of small size, first, the  DDS unit is designed to replace the Res unit in the backbone network. By connecting features of differ‐ ent levels across layers, complete and rich multi-scale features can be obtained, and small-size dam‐ age detection can be completed. Second, the gradient harmonized mechanism is introduced into the  classification loss function, and the weight of small-size damage is dynamically adjusted to make it  fully trained. Aiming at the low contrast between damage and background, first, the coordinate atten‐ tion mechanism is embedded in the deep network layer of the backbone network to enhance the  model´s attention to damage characteristics and reduce the interference of background noise. Second,  the accurate decoupled head is designed to improve detection accuracy by solving the contradiction be‐ tween classification and location requirements for features. 【Findings】Experimental results demon‐ strate that the mean average precision of this model is increased by 3.92% compared with that of YO‐ LOv4, and the detection accuracy of small-size crack damage and low-contrast wear damage is im‐ proved by 4.32% and 4.24%, respectively, which effectively solves the problems of missed detection  and false detection.