Research On Key Techniques Of Sub-pixel Registration Algorithm For Marine Remote Sensing Image

Offshore marine environment,marine disasters and maritime emergencies usually have rapid and dynamic characteristics(rapid changes in the hour),while the daily observation of the sun synchronous orbit(polar orbit)ocean satellites is difficult to meet the needs of daily monitoring.The geostationary satellites can use the same remote sensor to continuously observe the same area of interest,which is the best way to carry out high frequency ground observation.However,any satellite platform is subject to flutter interference,because the geostationary satellite imaging system integration time is generally longer,these flutter will seriously affect the remote sensing image quality.At present the satellite platform flutter measurement calculation and suppression technology have developed to a higher level.Satellite platform suppression technology and related equipment can eliminate most of the platform flutter,but for low-frequency satellite platform posture drift is helpless.The orbital height of the geostationary oceanic imaging radiometer is 35800 km.The detection range is 8 bands of visible light to near infrared.The angular resolution is 7?rad,and the ground resolution(star point)is 250 m.The 2048 × 2048 silicon CMOS surface array detection Device LUPA4000.Because of the high orbital height of the geostationary ocean imaging radiometer,the seismic energy of the ocean is weak and the ground resolution is high,it is particularly sensitive to the attitude drift of the satellite platform during the imaging process.In order to improve the signal-to-noise ratio of the system,multiple accumulation methods can be used.Static satellite marine imaging radiometer visible module up to 16 times cumulative.In the process of accumulation,the satellite platform there is a low frequency of attitude drift.According to the data of the satellite platform,the stability of the satellite platform is 5 × 10-4 ?,so the image is shifted by up to 1.8 pixels during the 16 th accumulation process.If you do not deal with the direct accumulation,will inevitably lead to image blur,affecting the image quality.In this paper,the sub-pixel registration algorithm of remote sensing image is proposed for the first time according to the idea of zeroing.First,the image is divided into different sub-images according to the optimal criterion.Secondly,the image is segmented by Ostu's Canny algorithm,The extracted image uses the SURF algorithm to extract the feature points,and finally windowing around the key points,the window size is 200 × 200 pixels.The matrix sub-pixel offset is calculated using the matrix multiplication phase correlation method in the window.The sub-pixel offset of the entire image is finally obtained by integrating the offset of all the sub-images.Based on the simulation,in order to improve the processing speed of the algorithm after hardware,this paper proposes an improved Shannon entropy low information feature point elimination algorithm and improved SURF(Fig.8).In this paper,Algorithm: Improved Shannon entropy low information feature point elimination algorithm reduces the number of feature points involved in matching and improves the execution speed of the algorithm.The improved SURF algorithm reduces the dimension of the eigenvector descriptor from the original 64-dimension to 36-dimensional,which obviously improves the feature point matching speed,and the improved SURF algorithm can be processed in parallel.These improvements will greatly enhance the hardware execution speed of the algorithm.The research content and innovation of this paper have the following four aspects:1)As the remote sensing image size 2048 × 2048,then the process of storage resources and computing resources will be very large,the general hardware such as FPGA and DSP will not be able to deal with,so this paper proposed remote sensing image sub-pixel registration algorithm pairs Image parallel processing to improve the speed of the algorithm at the same time also solve the FPGA,DSP and other hardware can not handle large size remote sensing image problems.And because the remote sensing image sub-pixel registration algorithm based on matrix multiplication phase correlation method,the algorithm has a significant inhibitory effect on the noise,so even if the remote sensing image noise,the algorithm can still get a higher precision sub-pixel offset Estimated value.2)Improved Shannon entropy low information feature point elimination: The algorithm not only greatly reduces the number of feature points involved in matching,but also improves the correctness of matching.So you can significantly improve the algorithm's computing speed.The SURF algorithm can obtain many feature points.However,there are many unmatched points in the process of image matching.In this paper,we propose a low eigenvalue feature point elimination algorithm to improve Shannon entropy.The improvement is mainly reflected in the fact that not only the discrete pixel values of the feature regions are considered,but also the interrelationship between the feature center point and the surrounding other pixels are also taken into account.3)The improved SURF algorithm realizes the parallel computation of the main direction calculation and the eigenvector descriptor,and also reduces the feature vector descriptor from the original 64-dimension to 36-dimensional.These improvements can not only improve the execution speed of the algorithm,But also improve the correct rate of matching: SURF algorithm is mainly reflected in the gradient of the feature points using radial gradient instead,this alternative can achieve the character description of the rotation invariant.In the process of generating the eigenvector descriptor,the traditional SURF algorithm uses a square region with a region size of 20 S × 20S(S is the scale of the scale space where the feature points are located).Instead,the improved SURF algorithm uses a circular area with a radius of 20 S,and the coordinate system rotation step is omitted.The circular region of 20 S is divided into nine characteristic sub-regions.Each sub-region is described by four features,so that a 36-dimensional eigenvector descriptor is generated,which greatly reduces the dimension of the descriptor.Based on the detailed study and experimental verification of the subpixel registration algorithm of remote sensing image,the hardware architecture of the improved feature segment extraction algorithm of the SURF algorithm is introduced in detail.The hardware architecture of matrix multiplication phase correlation method for subpixel offset estimation is introduced.hardware Architecture Based on Regression Learning Image Interpolation Algorithm.Based on the deep study of the principle and steps of the algorithm,the algorithm is applied to the hardware hardware implementation segmentation,which provides the foundation and guidance for the hardware implementation of the sub-pixel registration algorithm of the remote sensing image.