基于轻量化快速卷积与双向加权特征融合网络的船舶裂纹检测

Ship crack detection method based on lightweight fast convolution and bidirectional weighted feature fusion network

[目的]针对人工目视与超声波方法的船舶裂纹检测存在效率低下、成本高昂和危险性高的特点,提出一种基于深度学习的船舶裂纹检测方法。[方法]首先,在原模型YOLOv5s的主干网络中使用轻量化卷积结构(GSConv)替代标准卷积并融入注意力机制,在降低主干网络参数量与计算量的同时,提升主干网络对裂纹特征的提取能力;然后,在网络的颈部使用基于PConv构建的C3_Faster替代原C3模块,提升模型的图像处理速度,增强模型快速性;最后,设计一种简化的双向加权特征融合网络(BiFFN)以改进原模型YOLOv5s中的特征聚合网络,提升裂纹的语义信息与位置信息的融合效果,以及模型对裂纹的识别准确度与定位精度。[结果]通过对船舶裂纹原始数据与增强数据的学习,所提改进模型实现了94.11%的检测精度和93.50%的召回率,模型的计算量降低了17.93%,参数量降低了15.81%。[结论]研究表明,基于轻量化快速卷积与双向加权特征融合网络(MLF-YOLO)的船舶裂纹检测方法,实现了模型轻量化与较高的检测精度和召回率,结果可为开发自主无人机船舶检测提供参考。

[Objective]As traditional ship crack detection methods based on artificial visual inspection and ultrasonic methods in ship repair and inspection processes have the characteristics of low efficiency,high cost and high danger,a ship crack detection method based on deep learning is proposed.[Methods]First,a lightweight convolutional structure(GSConv)is used to replace the standard convolution and introduce an attention mechanism in the backbone of YOLOv5s to achieve the reduction of network parameters and computational complexity while enhancing the ability to extract crack features.Second,C3_Faster constructed by a fast convolutional structure is used instead of the original C3 module in the neck of the network to improve the processing speed of the model and enhance its rapidity.Finally,an improved bidirectional weighted feature fusion network(BiFFN)is used to enrich the semantic and positional information of cracks in the feature map and improve the model's crack identification accuracy and location precision.[Results]By training on both original and augmented ship crack datasets,the proposed method achieves a detection accuracy of over 94.11%and a recall rate of over 93.50%,while reducing the computational complexity by 17.93%and parameter count by 15.81%.[Conclusion]This study demonstrates that the proposed ship crack detection method based on GSConv and BiFFN achieves lightweight model architecture and high detection accuracy and recall rates,providing useful references for the development of UAV/ship autonomous inspection systems.