基于秩分解和强语义信息融合的电力巡检算法

Power patrol inspection algorithm based on rank decomposition and strong semantic information fusion

针对现行电力巡检方法对于高似然目标区分能力较差、检测速度较慢等问题,提出TR-YOLOv5模型。在网络第0层引入卷积注意力机制模块(CBAM),加强网络对细粒度特征的提取能力,并在网络最深层借助Transformer注意力进行编码,加强语义信息的传递能力。对于模型残差结构中的3×3卷积进行秩分解,压缩模型的冗余参数量。在特征融合阶段提出GPAN结构,以GSPP控制各尺度的变换,提高特征融合对各尺度信息的融合。在主干网络与同尺度特征融合结构的连接中加强了语义信息的融合,提高模型的检测能力。在模型训练过程中,以边框回归损失函数(SIOU)和CrossEntropy Loss作为IOU和分类损失回归函数提高模型的定位、分类能力。将训练完成的模型采用PyQt进行封装,提高了人机交互体验。实验结果表明,TR-YOLOv5模型检测平均精度值(mAP)达到97.1%,模型浮点运算量减少到3.6 GFLOPs。消融实验与对比试验证明了TR-YOLOv5模型能有效解决电力巡检过程中的前述问题。

The TR-YOLOv5 model is proposed to address the problems of poor differentiation ability and slow detection speed of existing power patrol methods for high-likelihood targets. CBAM attention is introduced in layer 0 of the network to enhance the network′s ability to extract fine-grained features;and in the deepest layer of the network, encoding is performed with the help of Transformer attention to enhance the semantic information transfer capability. For the 3×3 convolution in the residual structure of the model, rank decomposition is performed to compress the amount of redundant parameters of the model. The GPAN structure is proposed in the feature fusion stage to control the transformation of each scale with GSPP to improve the fusion of feature fusion to information at each scale. The connection of the backbone network with the same-scale feature fusion structure is used to enhance the fusion of semantic information and improve the detection capability of the model. In the model training process, SIOU and CrossEntropy Loss are used as IOU and classification loss regression functions to improve the localisation and classification ability of the model. The completed training model was wrapped in PyQt to improve the human-computer interaction experience. The experimental results show that the average accuracy(mAP) of the TR-YOLOv5 model detection reaches 97.1% and the model floating point operations are reduced to 3.6 GFLOPs. ablation experiments and comparison tests demonstrate that the TR-YOLOv5 model can effectively solve the aforementioned problems in the power inspection process.