针对空中鸟群特征的全维动态卷积识别算法

Full-dimensional Dynamic Convolution Recognition Algorithm for Aerial Bird Flock Features

机场驱鸟应用中,存在空中被驱离鸟群距离远、特征不明显、不易识别等难题,由此提出一种针对空中鸟群特征的全维动态卷积识别算法。该算法采用动态K值的K-Means++算法对数据集中的目标样本进行聚类,获取更符合不同目标尺度的锚框,提高多目标定位及其图像分割精度。在通用的YOLOv5s目标检测识别的骨干网络引入全维动态卷积模块,提取特征时动态卷积层会自动调整卷积核的大小和形状,使其适应不同鸟类的特征,通过动态卷积提取的鸟类特征使数据更具有代表性。针对输入图像经过多次卷积和池化操作之后产生的多个特征图,运用相干积分对不同特征图进行分离处理,筛选剪去特征差异性不明显的特征通道,从而减少需要计算的信息量,优化YOLOv5s算法检测精度和计算复杂度,实现识别网络轻量化,解决空中鸟群等小目标特征难提取的问题,提高鸟类识别准确率。

Aiming at the problem of long distance from the flock of birds in the air,inconspicuous features and difficult identification in the application of airport bird repelling,a full-dimensional dynamic convolution recognition algorithm for the characteristics of birds in the air is proposed,which uses the dynamic K-value detection K-Means++algorithm to cluster the target samples in the data set,obtain anchor frames that are more in line with different target scales,and improve the accuracy of multi-target localization and image segmentation.The full-di⁃mensional dynamic convolution module is introduced into the backbone network of general YOLOv5s object detection and recognition,and the dynamic convolutional layer automatically adjusts the size and shape of the convolution Kernel when extracting features to adapt it to the char⁃acteristics of different birds,and makes the data more representative by dynamically convolutioning the extracted bird features.Aiming at the multiple feature maps generated by the input image after multiple convolution and pooling operations,the coherent integration is used to sepa⁃rate different feature maps and screen and cut off the feature channels with less obvious feature differences,thereby reducing the amount of in⁃formation that needs to be calculated,thereby optimizing the detection accuracy and computational complexity of the YOLOv5s algorithm,re⁃ducing the amount of information that needs to be calculated,realizing the lightweight identification network,solving the problem that it is dif⁃ficult to extract small target features such as aerial bird flocks,and improving the accuracy of bird identification.