双池化及双层注意力融合的杆塔鸟巢识别算法

A tower bird nest recognition algorithm based on double pooling and double layers attention fusion

针对复杂环境下输电杆塔鸟巢识别率较低的问题,提出基于YOLOX的双池化及双层注意力融合算法。在特征提取方面,通过使用双池化优化空间金字塔池化结构,减少空间金字塔池化的信息丢失,提升信息在模型中的传递效率;在特征融合方面,提出基于双层特征的注意力融合方式对特征进行融合与加强,提升了模型对鸟巢复杂特征的表征能力;同时将RepVGG模块引入到特征融合和检测头部分,在不增加推理速度的同时提升模型的编码能力,以增强图像中鸟巢的识别能力;引入CIoU函数优化损失函数,改善模型的定位精度。在自制鸟巢数据集的实验结果表明,改进算法的精度均值可达到94.3%,相较于原始YOLOXs模型提高了3.7%,检测速度为46帧/s,可识别在遮挡、光照强弱等复杂环境下的目标,满足输电杆塔巡检中对鸟巢识别的性能需求。

For the problem of low recognition rate of bird nests in transmission towers under complex environment,the YOLOX-based double pooling and double layers attention fusion algorithms were proposed.In terms of feature extraction,the spatial pyramid pooling structure was optimized by using double pooling to reduce the information loss of spatial pyramid pooling and improve the information transfer efficiency in the model.In terms of feature fusion,the attention fusion method based on double-layer features was proposed to fuse and strengthen the features,which improved the model's ability to characterize the complex features of bird nests.Meanwhile,the RepVGG module was introduced into the feature fusion and detection head part to enhance the encoding ability of the model without increasing the inference speed to enhance the recognition of bird nests in images.The CIoU function was introduced to optimize the loss function to improve the localization accuracy of the model.The experimental results on the homemade bird nest dataset show that the mean average accuracy of the improved algorithm can reach 94.3%,which is 3.7%better than the original YOLOXs model,and the detection speed is 46 frames per second,which can identify bird nests in complex environments such as occlusion and light intensity,and meet the performance requirements for bird nest identification in transmission tower inspection.