通信与信息技术

提出一种基于深度学习技术的发动机钢管内壁缺陷检测方法，旨在提高检测的准确性和效率。传统的检测手段依赖人工观察，容易受主观和环境因素影响，准确性有限。针对这些局限性，设计了一种改进的网络结构，结合ConvNeXt V2主干网络和AFPN(Adaptive Feature Pooling Network)颈部，以增强特征提取和多尺度目标检测的能力。ConvNeXt V2作为主干网络，采用创新的卷积结构和全局响应归一化（GRN）技术，有效提升了特征表达能力，同时降低了计算复杂度。GRN技术通过聚合和归一化操作，增强了通道间的对比度和选择性，防止了特征崩溃现象，提高了特征的多样性。AFPN颈部则利用自适应特征池化技术整合不同尺度的特征图，增强了模型对多尺度目标的检测能力，尤其对复杂场景中缺陷的识别至关重要。改进的网络结构通过特征融合技术，提高了模型对不同类型缺陷特征的敏感度。结果表明，使用改进后的网络mAP50%提高了8.9%。

In this paper, a defect detection method for the inner wall of engine steel tube based on deep learning technology is proposed, which aims to improve the accuracy and efficiency of detection. Traditional detection methods rely on manual observation, which is easily affected by subjective and environmental factors, and has limited accuracy. In view of these limitations, an improved network structure is designed, which combines the ConvNeXt V2 backbone network and the Adaptive Feature Pooling Network(AFPN) neck to enhance the ability of feature extraction and multi-scale object detection. As the backbone network, ConvNeXt V2 adopts innovative convolution structure and Global Response Normalization(GRN) technology, which effectively improves the feature expression ability and reduces the computational complexity. GRN technology enhances the contrast and selectivity between channels through aggregation and normalization operations, prevents feature collapse, and improves feature diversity. In the neck of AFPN, the adaptive feature pooling technology is used to integrate the feature maps of different scales, which enhances the model's ability to detect multi-scale targets, especially for the identification of defects in complex scenes. The improved network structure improves the sensitivity of the model to different types of defect features through feature fusion technology. The results show that the use of the improved network mAP 50% is increased by 8.9%.