Research And Implementation Of Intelligent Detection Algorithm For Remote Abnormal Image

With the rapid development of deep space exploration,the number of remote sensing satellites is increasing.The rich image data brought by the increase of satellites has been widely used in many fields,such as remote sensing mapping,marine production,geological exploration,military security,education and communications.These data play an important role in these areas.The application of remote sensing data has become an important means for humans to obtain information.However,due to the effects of sensor imaging system anomalies,information transmission media interference,and space particle radiation,remote sensing data will inevitably be disturbed by various noise signals during the reception,transmission,and recording,which will cause image anomalies.At this stage,remote sensing anomaly image detection systems use methods such as fast-view image method and code stream parity method.These methods have the disadvantages of low detection accuracy and large resource consumption.It is difficult to achieve autonomous intelligence in the face of increasing anomaly detection.Therefore,by analyzing the causes of abnormal images,this thesis proposes an intelligent detection algorithm for remote sensing abnormal images to solve the problem of detecting abnormal images.There are the following reasons for the abnormal remote sensing images: First,due to the high-energy particles in space,the abnormality of the camera’s photosensitive elements cause local abnormal image;Secondly,the space storage element is subject to particle bombardment or the code stream data is disturbed during transmission,which will cause the image code stream data to randomly flip and generate non-local abnormal image.In order to detect the above abnormal situation and remove the influence of abnormal images on subsequent applications,this thesis proposes an abnormal image detection algorithm based on remote sensing denoising network.The algorithm judges whether the image is abnormal by comparing the correlation between input and output of network.This algorithm can not only remove abnormal areas in the image,but also retain the feature information in the image.The thesis discusses remote sensing image preprocessing,abnormal scene simulation,noise feature extraction,neural network design,training method improvement,algorithm of threshold evaluation and does the following research:(1)Aiming at the problem of insufficient anomaly datasets for network training,the thesis designs an algorithm to transform remote sensing data into anomalous datasets.First,the thesis uses multi-scale spatial feature method and gradient fusion method to design image preprocessing algorithm to filter out pure remote sensing images from a large number of remote sensing images.Then,according to the causes of remote sensing anomalies,thesis simulate abnormal transmission scenes of remote sensing,and then obtain abnormal images Data set;Finally,for the abnormal image data,we calculate mathematical distribution rules of the abnormal features,and the characteristics of the abnormal distribution are fitted with formulas.(2)Aiming at the insufficient computing power and redundancy network design,this thesis designs a pixel-level convolution model and codec network structure,which ensure the algorithm can run deeper and more complex networks with the same computing power,and improve the network expression ability.Aiming at damaging image detail information,this thesis designs a gradient-based loss function to improve the network detail preservation ability and network division effect.(3)Aiming at data enhancement in experiments,the thesis proposes a pre-training method based on the anomaly distribution law.This method uses a noise distribution function to generate a large number of fake abnormal images,and uses them to pre-train the network.The data generated by this method is of high quality and large quantity,which can promote the network to converge within a specific target weight range.Next,the network is initialized by pre-training weights,and then the abnormal images generated by the simulation of abnormal transmission process are used to fine-tune the network to make up for the shortcomings of insufficient abnormal data,and this method improves network performance.(4)Based on the analysis of papers in threshold selection,this thesis proposes a threshold selection algorithm based on data analysis.This method can select the most appropriate judgment threshold in the statistical detection data.and,the algorithm can evaluate the network design according to the algorithm’s distinguishing degree.