Research On Super-resolution Recognition Method Of Infrared Target Based On Machine Learning

Infrared imaging technology is important for target detection and recognition beacuase of its concealment,day and night visibility,and reflecting temperature characteristics.With the development of science and technology,spatial resolution and spectral resolution of remote sensing detectors have been further improved.However,due to the infrared light diffraction limitation and high hardware cost,the resolution of infrared images is generally low,the noise is large,the edges of small infrared targets are blurred,and the texture is not clear,which brings great difficulties to the accurate detection and recognition of infrared targets.In view of the aforementioned problem,this thesis proposes a learning-based infrared remote sensing Super Resolution Object Recognition(SROR)algorithm.The software technology is used to improve the resolution of the infrared remote sensing image,and then the learning-based target recognition algorithm is used to detect and recognize reconstructed small targets.In order to improve reconstruction results of infrared remote sensing images,a sensor downsampling simulation model is used to simulate the degradation process of infrared images,and degraded infrared data was used to train the super-resolution network WDSR.In order to improve the accuracy of small infrared target detection,Faster RCNN was improved and training was is performed using transfer learning.In order to improve the accuracy of infrared small target detection,the structure of Faster RCNN is improved,and transfer learning is used for training.The improved network significantly improves the detection effect of small targets.Through the combination of super-resolution reconstruction and target recognition network,the detection accuracy of the target reached 88.59%,and the recall rate reached 81.45%.Main content and innovations of the article are as follows:1)A super-resolution reconstruction algorithm for infrared salient regions based on sparse coding(Sr SR)is proposed.Based on the principle of sparse coding,this research improves the feature extraction operator,selectively performs super-resolution reconstruction on salient regions segmented by visual saliency,and filters non-salient regions with Gaussian interpolation.The experimental results show that salient area of the reconstructed image is enhanced in detail,and background noise is suppressed.In terms of efficiency,the reconstruction time of the background area is reduced,and the reconstruction speed is significantly improved.2)It is proposed to train WDSR with images processed by analog sensor sampling mechanism,ST-WDSR.Assuming that the high-resolution image is a real scene,this thesis uses a sensor downsampling simulation model including single sampling and oversampling to downsample the high-resolution image.The downsampled images are input to taining network instead of bicubic interpolation images.A quantitative evaluation method of target characteristics was is proposed.Based on the quantification of remote sensing,a quantitative evaluation method of target characteristics is proposed.Using the 8th band(center wavelength 0.862μm,resolution 10m)and 8A band(center wavelength 0.842μm,resolution 20m)images taken by Sentinel-2A MSI,the center wavelength of the image is similar and the resolution is different.Quantitative comparison and analysis of 8A image reconstruction results and 8 bands.The experimental results show that the geometric and radiation characteristics of targets in images reconstructed by ST-WDSR are closest to the 8-band image,and the degree of approximation is increased by 8.9% and 14% compared to WDSR.The results are in favour of the subsequent target detection and recognition.3)Using single sampling and oversampling methods to change the resolution of the infrared image,and study the change law of target characteristics.The sensor sampling simulation model is used to sample high-resolution target(car and ship)images,and the target’s radiation characteristics,geometric characteristics,and Hu invariant moment characteristics are counted.Experiments have confirmed that the geometric characteristics of automobile targets begin to change back when the resolution is lower than 0.7m.The radiation characteristics change slowly and the mean value decreases.For large ship targets(100 meters in length),the geometric characteristics change more than 10% when the resolution is lower than 12 m.When the resolution is lower than 20 m,it is difficult to determine the specific shape of the ship.Geometric changes tend to point light sources.Radiation characteristics change significantly,and both the mean and standard deviation show a downward trend.The scale retention of M1 and M2 invariant moments of Hu invariant moments is better than that of M3 moments.Hu moments change significantly when the resolution is lower than 13 m.4)An infrared remote sensing super-resolution target recognition algorithm(SROR)is proposed.The model is mainly composed of two parts.The first part is a pre-processing module,which adopts the structure of “segmentation—super segmentation—segmentation” to crop and super-resolution reconstruct images.The second part is the target recognition module.In order to improve the detection effect of Faster RCNN on small targets,the low-level features in the residual network are also used as shared features of RPN to avoid small targets due to the expansion of the receptive field and the dimensionality reduction of the pooling layer The influence of features has improved the scale invariance of the detection network.The soft-NMS method is used instead of non-maximum suppression,and the confidence threshold is continuously updated,which avoids the problem of deleting overlapping boxes of the aggregation target as repeated boxes.Experiments show that SROR can accurately detect small infrared targets within 10 pixels.Compared with Faster r-cnn trained by infrared target,the accuracy of Faster R-CNN and the recall rate are increased by 5.33% and 12.22% respectively.Note that the recall rate of small(less than 20 pixels)target recognition is higher than that of Faster R-CNN by 13.25%.