Research On Airborne Infrared Scene Classification Algorithm Based On Deep Learning

This paper takes airborne infrared early warning detection as the research background.Different complexity of the scene will bring different effects on infrared target detection and tracking.This leads to the general problems of current target detection and tracking algorithms,such as poor adaptive ability,susceptibility to interference from scene factors,and high false alarm rate.Therefore,this paper focuses on the research of scene classification algorithms,utilizing semantic segmentation pixel-level classification methods to achieve multi-scene category division and obtaining different region segmentation results from infrared images.The content of this research mainly includes the following aspects:(1)Firstly,an airborne infrared image semantic segmentation dataset is established.In addition,an image compression method based on residual map enhancement mapping is proposed,which not only realizes the compression of wide dynamic gray range,but also enhances the detailed information on the infrared image.Therefore,it provides a reliable dataset for subsequent segmentation model training.(2)From the perspective of the lightweight model,the lightweight Ghost module is used to improve the BiSeNetV1 and BiSeNetV2.Ghost-based lightweight bilateral infrared semantic segmentation models have been proposed in the research,namely GIR-BiSeNetV1 and GIR-BiSeNetV2,which greatly reduce the number of model parameters at the expense of small real-time performance.(3)On the basis of GIR-BiSeNetV2,a pixel-level infrared scene classification algorithm based on AIR-BAESe Net segmentation network is proposed.In this algorithm,the lightweight atrous convolution pyramid module is combined with the spatial attention mechanism for the first time,which effectively enhances the ability of multi-scale spatial detail feature extraction.Meanwhile,the Ghost bottleneck layer based on the channel attention mechanism is designed,which can effectively suppress the influence of the common redundant features in the channel dimension.The experimental results prove that the algorithm has great advantages in solving the problem of classification confusion and fuzzy boundary segmentation,and achieves efficient pixel-level infrared scene classification with fewer model parameters.