Application Of Image Dehazing Algorithm Based On Frequency Domain And Data-driven In Crop Monitoring

The use of modern science and technology,including low altitude remote sensing through unmanned aerial vehicles(UAVs),has greatly improved agriculture by allowing for accurate monitoring of crop growth.However,these dynamic monitoring methods can be affected by natural weather events,particularly rain and haze.Haze in natural images often has nonhomogeneous features such as filaments,masses,and mist,with high-frequency parts containing variable background textures and haze shapes,and low-frequency regions dominated by uniform information.While current methods utilizing convolutional neural networks have made significant progress in single image dehazing,they often neglect the intrinsic patterns present in hazy images.To address this issue,this thesis proposes a frequency division dehazing neural network that utilizes prior knowledge of hazy images to enhance and dehaze UAV monitoring images of crops.The proposed network processes shallow feature maps through high-,medium-,and low-frequency branches,allowing for a flexible architecture in which the lowerfrequency branch is less redundant due to the simpler background and haze shapes it handles.By integrating knowledge from all branches through feature fusion,the proposed network is able to fully utilize the various frequency characteristics present in hazy images.Experiments on synthetic and real hazy images have shown good performance,with PSNR and SSIM reaching 39.51 and 0.9931 respectively on the traditional synthetic dehazing dataset,demonstrating the superiority of the proposed network over several stateof-the-art methods,as well as the effectiveness of using prior knowledge in hazy images.And by verifying on synthetic and real agricultural UAV hazy images,it proves that our method is also suitable for crop monitoring field,which can provide high-quality images for subsequent digital agriculture tasks such as crop growth monitoring and yield estimation.