Airborne photoelectric video target intelligent detection and identification method

Abstract

The invention discloses an airborne photoelectric video target intelligent detection and identification method. On the basis of a YOLOv3 model, rectangular convolution is adopted to extract strip-shaped target features such as a bridge, expansion convolution is adopted to expand a receptive field and reserve spatial structure information of a multi-scale target, a visual attention mechanism is introduced into a sampling branch on a feature pyramid to endow the model with different weights of target features of different regions and different channels, and a convolution mode of a residual module is improved into depth separable convolution to reduce calculation complexity. The method has the following advantages: high aerial photography target detection precision is kept, and meanwhile, high aerial photography target detection and recognition speed can be achieved on an airborne embedded system.

Description

technical field [0001] The present invention relates to the technical field of airborne photoelectric radar detection and recognition, in particular to a real-time detection and recognition method of visible light video targets based on an embedded system, which is suitable for air, sea, sea, Accurate and real-time detection and recognition of multiple types of target images. Background technique [0002] Real-time detection and recognition of aerial image targets based on embedded systems is one of the important research directions in the field of computer vision. It extracts and uses target feature information on embedded devices with limited computing resources, locates targets in images, and searches for them. Classification of multiple types of targets has broad application prospects in military target detection, aerial search and rescue, remote sensing image analysis and other fields. [0003] In the target detection and recognition of the aerial image of the airborne...

AI Technical Summary

Problems solved by technology

However, the YOLO method does not fully consider and utilize the target features in the aerial image; and the YOLO method will treat the features extracted in different regions of the image, different channels of the feature map, and different layers of the neural network equally, which will cause the model to be more subject to decision-making. Many redundant features interfere, making it difficult for the effective features of the target to be fully learned by the model

[0004] At the same time, the real-time performance of target detection is also a problem that must be considered in the detection and recognition of aerial images based on embedded systems. Although the deep learning target detection and recognition method has high accuracy, it also has high computational complexity and high computational resource overhead.

Even though YOLO belongs to the one-stage model, which is faster than the two-stage target model, it is difficult to meet the speed requirements for real-time detection and recognition of targets on the airborne embedded system.

The Tiny-YOLO series model uses a lightweight neural network with fewer layers to speed up the detection speed. Although it can detect targets in real time on the airborne embedded system, the large number of convolutional neural network layers reduces the feature extraction capabilities of the model. decrease, which in turn affects the accuracy of target detection and recognition

Images