Research On Convolutional Neural Network Image Compression Algorithm For UAV Remote Sensing Images

With the rapid development of sensor technology,the spatial,radiative and spectral resolution of UAV remote sensing images has been continuously improved,and the storage space occupied by a single image has become larger and larger.The huge amount of data brings challenges to the storage,transmission,management and application of images.High-efficiency image compression algorithms are crucial for image storage,transmission,management and even the development of UAV remote sensing technology.Although traditional image compression algorithms have achieved certain compression effects,traditional image compression methods have encountered bottlenecks in improving the quality of image reconstruction,while convolutional neural network image compression algorithms have made great progress in improving the quality of reconstructed images.In view of the characteristics of UAV remote sensing images with a large number of bands and a large amount of data,this paper proposes a convolutional neural network image compression algorithm for UAV remote sensing images.The main work of this paper is as follows:Aiming at RGB images of UAV remote sensing,this paper designs an end-to-end convolutional neural network image compression algorithm consisting of self-codec,quantization structure,entropy coding,rate-distortion optimization and other links.The autoencoder reduces the dimension of the image and reduces the size of the image to 1/64 of the original;the quantization structure converts the floating-point data output by the auto-encoder into an integer;entropy coding performs arithmetic coding on the quantized data,and estimates Code stream;rate-distortion optimization Jointly optimizes image code stream and loss.The experimental results show that the reconstructed image quality of the algorithm in this paper is better than that of JPEG and JPEG2000 compression algorithms,and the PSNR of reconstructed images is improved by about 3d B compared with JPEG compression algorithm and about 1.5d B compared with JPEG2000 compression algorithm.In view of the lack of practical application of the convolutional neural network image compression algorithm,this paper transplants the algorithm to the NVIDIA Jetson TX2 edge computing device to carry out engineering application experiments of UAV remote sensing image compression.The experimental results can meet the real-time collection of UAV remote sensing data.,compression and transmission requirements.Aiming at the problem of low reconstruction quality of remote sensing multispectral images due to various quantification types,this paper proposes a data preprocessing method.By converting the DN value of multispectral images into surface reflectance,it eliminates the problems caused by instruments and observation conditions when different sensors collect data.The standardization of multi-source remote sensing multi-spectral images is realized.Aiming at the problem of the large number of bands in multispectral images and the redundancy between spectral information,this paper proposes a convolutional neural network image compression algorithm based on an improved autoencoder.The three-dimensional convolutional structure is introduced into the autoencoder to improve the spectral information.The ability to express information features improves the reconstruction quality of multispectral remote sensing images.The experimental results show that the quality of the reconstructed images of this algorithm is better than that of the JPEG compression algorithm,and the peak signal-to-noise ratio is 2 d B higher than that of the JPEG compression algorithm,which avoids the reconstruction of the JPEG compression algorithm.There are block effects and missing texture features in the image.The multispectral image of GF-6 satellite has the characteristics of large width,special quantization type,and large number of bands.In order to verify the robustness of the multispectral image compression algorithm of the improved autoencoder convolutional neural network in this paper and the applicability between different data sources The improved multispectral image compression algorithm is used to process the multispectral images of the GF-6 satellite.The experimental results show that the algorithm in this paper can restore the features of the ground objects well,and verified by the quantitative remote sensing experiment,the multi-spectral reconstruction image of the algorithm can meet the needs of the quantitative remote sensing experiment.