| As a new battlefield reconnaissance equipment in recent years,UAVs have become a mainstream equipment in the modern battlefield due to their high flexibility,stealth and low cost.However,when using a single visible light sensor on the UAV’s airborne platform for reconnaissance,the acquired aerial images will be interfered by the smoke and dust in the battlefield environment,leading to the problem of target loss.Therefore,it is of great importance to study target detection for UAVs in the smoke and dust environment.The main research elements of the thesis are as follows:First,to address the problem of blurred information and missing details in target images detected using a single visible sensor in a sooty environment,images are acquired using visible and near-infrared sensors and fused to reduce the interference of soot.A fusion soot removal algorithm based on NSST decomposition is designed for YUV colour space and NSST decomposition,using guided filtering for the decomposed high-frequency components and a fusion strategy with adjacent grey values being larger to highlight the detail information in the high-frequency components,and an SSR algorithm for the decomposed low-frequency components to do enhancement processing and a fusion strategy based on region energy to highlight the background information in the low-frequency components.The YUV colour space is combined to reduce the distortion of the image colours during processing.Finally,a comparison with common defogging algorithms is performed to verify the feasibility of the algorithm.Secondly,the improvement of the SSD model is based on the small target and complex background in the aerial image model.For the problem of lack of semantic,detailed information in shallow feature maps in the SSD model,Designed a feature fusion mechanism,Enrich the semantic and detailed information of the shallow feature graph by adding the feature layer and the feature layer with the semantic information obtained from the recursive reverse path;For the inadequacy of SSD models to focus on channels as well as spatial information,A hybrid attention module combining channel and space is introduced to improve the overall attention ability of the model;For the problem of the mismatch of the prior box to the smallscale target in the SSD model,The proportion of the a priori box was adjusted;After improving the SSD model based on the above strategy,Was aligned with the other models on the homemade dataset,The advancement of the improved model is verified.Finally,in order to integrate the fused smoke removal section with the target detection section,the PYQT software was used to create a host computer to verify the functionality of the fused smoke removal and target detection sections and to complete the commissioning of the overall system linkage. |
