Lightweight air-to-ground target detection method based on attention gradient

Abstract

This invention discloses a lightweight air-to-ground target detection method based on attention gradients, belonging to the field of air-to-ground target detection. The implementation method of this invention is as follows: An attention gradient feature extraction network is constructed, mainly composed of a CBL module (convolution + batch normalization + ReLU activation), a linear bottleneck structure, and an attention gradient module, to extract original image features and improve the network's ability to represent small targets in air-to-ground images; a feature pyramid network is used for feature fusion to improve the detection accuracy of targets at different scales in air-to-ground images; a dynamic positive and negative sample allocation strategy based on mathematical statistics is adopted to improve the efficiency of anchor box allocation; the convergence speed of the air-to-ground target detection network training is improved by decoupling the classification prediction module and the location prediction module; and end-to-end training of the model is achieved using cumulative loss calculation, improving the accuracy and efficiency of air-to-ground target detection, reducing model parameters, and making it easier to deploy on air-to-ground platforms with limited computing power.

Images