Unmanned aerial vehicle multi-scale target detection and identification method

Abstract

The invention discloses an unmanned aerial vehicle multi-scale target detection and recognition method, which is high in recognition rate, small in calculation amount and high in robustness. The method is realized through the following technical scheme: forming a real-time target detection and identification network model by adopting three parts, namely a trunk structure for extracting different scale features, a network neck (Neck) and a network detection head for predicting target information, performing feature extraction on an input image by adopting a CSPDarknet53 trunk network in a target detection and identification network-YOLOv4, and expanding the original three-scale feature output into four-scale feature output; using an improved two-branch PANet to reduce the number of convolution layers through which features pass; predicting a conditional probability value for each category by each detection output, directly obtaining a prediction result from the picture, and obtaining target information; and transmitting the four feature maps with different sizes to a detection head for joint training, and performing category judgment and position regression on the unmanned aerial vehicle target to obtain a detection and recognition result.

Description

technical field [0001] The invention relates to a real-time target detection and recognition model based on an improved PANet structure in the field of target detection, and is suitable for a multi-scale target detection method in scenes where image target scales are widely distributed and small target detection is difficult. Background technique [0002] Object detection and recognition is an important research topic in the field of computer vision. With the wide application of artificial intelligence technology in the field of computer vision, target detection as one of the representative problems in computer vision has attracted more and more attention, and target detection and recognition methods have been vigorously developed, especially in a variety of complex image processing. field. Among them, a variety of remote sensing image classification technologies based on deep learning are widely used in environmental monitoring, urban planning, disaster control, and agricu...

AI Technical Summary

Problems solved by technology

[0010] The purpose of the present invention is to address the shortcomings of the existing technology and the target scale distribution range is wide, and the detection of small targets is difficult. The UAV target detection and recognition method has a wide target scale distribution range and small target detection difficulty. The problem is to provide a UAV multi-scale target detection and recognition method with high recognition rate, small amount of calculation, strong robustness, and higher detection accuracy on the target data set.

Images