Full-dimensional Dynamic Convolution Recognition Algorithm for Aerial Bird Flock Features
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.
aerial bird flockYOLOv5s target detectionfull-dimensional dynamic convolutioncoherent integral
万方数据
维普
