基于卷积注意力的输电线路防震锤检测识别

Convolutional attention mechanism based object detection method for vibration damper

摘要:

为提高单目标多分类(Single Shot MultiBox Detector,SSD)网络模型对输电线防震锤的识别准确率，提出一种融合卷积注意力机制和SSD模型相结合的新方法。该算法采用残差网络ResNet作为骨干网络，引入卷积注意力机制将通道和空间注意力结合，通过压缩提取中间特征和利用权重系数更好地分辨出前景与背景，提高对输电线路中防震锤检测的精度和速度。训练时引入迁移学习策略，克服了模型训练困难问题。实验结果表明，提出的算法不仅提高了检测准确率，计算效率亦得到了提升。与经典SSD算法相比，输电线路的防震锤检测准确率提升了2.5%，检测速度达到了12fps识别效果明显提升,证明了新算法的有效性。

Abstract:

To improve the detection accuracy of (SSD) model for vibration damper, an attention mechanism based detection method is proposed. The method adopts ResNet as the backbone network instead of the VGG network, and introduces attention mechanism to improve the accuracy and speed of detection of vibration damper in transmission lines by extracting intermediate features through compression and using the weight coefficient to better distinguish the foreground and background. The introduced fused convolutional attention mechanism combines channel and spatial attention, and the performance jump is relatively obvious, while the computational efficiency is improved. A migration learning strategy is introduced to overcome the problem of difficult model training. The experimental results show that the SSD detection network model using the ResNet residual structure as the backbone and the fused convolutional attention mechanism improves the accuracy of seismic hammer detection in transmission lines by 2.5 percentage points,and completes vibration damper detection at 12fps. The recognition effect is significantly improved, which proves the effectiveness of the new algorithm.