Lightweight rotating target detection and recognition method based on airborne photoelectric video

Abstract

The invention discloses a lightweight rotating target detection and recognition method based on an airborne photoelectric video, and the method comprises a process of constructing a lightweight rotating target detection and recognition model, and the lightweight rotating target detection and recognition model carries out the feature extraction of a photoelectric video image through a feature extraction network improved by a channel splitting aggregation structure. According to the method, the lightweight deep neural network model can be used for quickly detecting and identifying the multi-type, multi-scale and multi-direction rotating targets in the airborne photoelectric video image, the detection and identification precision and stability are high, and the calculation complexity is relatively low.

Description

technical field [0001] The invention relates to the technical field of airborne photoelectric radar target detection and recognition. Background technique [0002] At present, there are some methods for detecting rotating objects in aerial video images through deep learning. Lightweight improvements are made, but the massively stacked pair efficiency in depthwise separable convolutions creates significant limitations, especially for 1×1 point-to-point convolutions. [0003] Some existing technologies have been improved, such as Tiny-YOLO, which uses a network with fewer convolutional layers to improve computational efficiency and speed, but with the reduction in the number of network layers, its feature learning ability is weakened, resulting in low detection and recognition accuracy . Therefore, a deep neural network model that can maintain high target detection and recognition accuracy and process photoelectric video targets in real time on an airborne embedded platform ...

AI Technical Summary

Problems solved by technology

[0004] On the other hand, for the detection and recognition of rotating targets in video images, although existing methods use methods such as regression of rotation angles (five-parameter method) or four-vertex regression (eight-parameter method) to represent and predict the rotation in aerial photoelectric video images target, but these methods usually also have the problem of low positioning accuracy of rotating targets

For example, the five-parameter method based on angle regression can only predict rectangles rotated at a specific angle, and it is difficult to accurately locate targets of more diverse shapes, and there are periodic problems in angle regression; although the eight-parameter method based on vertex regression can represent arbitrary shapes quadrilateral bounding box, but there are too many degrees of freedom in the shape of the bounding box, which will cause the deep neural network to be difficult to predict stably and have ambiguity in vertex ordering, which restricts the improvement of target detection and positioning accuracy

[0005] Summarizing the above existing technologies, it can be seen that whether they are angle-based or vertex-based methods in the detection of rotating objects, it is difficult to learn the object shape stably and accurately. However, the computational complexity of the product still needs to be greatly improved.

Images