Treffer: 基于多特征融合的集装箱船导轨缺陷检测算法.

In response to the issues of traditional container ship guide rail defect detection methods, which rely entirely on manual visual inspection and suffer from low efficiency and heavy workloads, a defect detection algorithm is proposed for container ship guide rails based on multi-feature fusion. A processing method of data adaptive resampling is designed to mitigate the impact of uneven distribution of defect types. A multi-gradient receptive field aggregation module is incorporated into the backbone network to aggregate features of guide rails with varying degrees of damage and their surrounding environmental features. Building on this approach, a hybrid attention mechanism is embedded after the residual analysis module to effectively guide multi-scale feature flows toward key feature information. At the feature concatenation points of the network, a feature reorganization upsampling operator is fused to expand the local receptive field of incoming features and effectively integrate global subtle feature information. Validation on the test set and comparison with manual efficiency demonstrate that the proposed improved algorithm achieves a mean average precision of 97.0% for guide rail defect detection, which is 2. 9 percentage points higher than the original YOLOv5 algorithm, effectively enhancing the detection accuracy of container ship guide rail defects. [ABSTRACT FROM AUTHOR]

针对传统集装箱船导轨缺陷检测方法完全依赖人工目视检查, 存在效率低、工作量大等问题, 提出一种基于多特征融合的集装箱船导轨缺陷检测算法, 设计了数据自适应重采样处理方法, 降低缺陷种类分布不均的影响, 在骨干网络设置多梯度感受野聚合模块, 聚合导轨不同程度破损特征和周围环境特征, 根据上述方法, 在残差分析模块后嵌入混合注意力机制, 有效引导多尺度特征流关注重点特征信息, 在网络的特征拼接处融合特征重组上采样算子, 扩张流入特征的局部感受野, 有效整合全局细微特征信息, 在测试集上的验证以及与人工效率的比对表明: 所提改进算法对导轨缺陷检测的均值平均精度可达到 97.0%, 相较原 YOLOv5 算法提升 2.9 个百分点, 有效提升了集装箱船导轨缺陷检测精度。 [ABSTRACT FROM AUTHOR]