Research On Technologies Of Infrared Small Target Detection Based On Multimodal Features

In military field,target detection is the key to the combat effectiveness of weapon systems such as early warning,interception,reconnaissance and long-range strike.Target detection technology based on optical features has the advantages of high recognition probability,fast recognition speed and good crypticity.At the same time,it also has broad application prospects and practical value in civil fields such as intelligent transportation and security.It is currently a research hotspot in the above fields.In recent years,it poses a major threat to China’s daily air security and wartime air defense and anti-missile with frequent occurrence of "Low,Slow and Small" unknown air information in key areas and the long-range cross-border reconnaissance and harassment of enemy aircrafts in border areas,as well as the development of target stealth and ballistic missile penetration technology.High accuracy detection of small targets under the condition of far distance identification has become a difficult problem to be solved urgently.Therefore,the intelligent detection and identification technology of infrared small targets based on multimodal features is studied in this dissertation,which is used to solve the problem of accurate detection of long-range targets in fulltime and provide accurate target detection information for weapon systems.It has farreaching influence and overall strategic significance on improving and optimizing the performance of the combat platform.In order to deal with the limitations of sensor detection ability and poor description of shape,structure and texture information of small targets,low signal to noise ratio,low radiation characteristics accuracy,poor infrared measurement feature extraction ability and other attributes,and to address the limitations of single modal data on small target feature description,this dissertation takes the optical imaging characteristics,infrared radiation characteristics and motion characteristics of the small targets as the detection characteristics to carry out the research on infrared small target detection technology based on multimodal features.The complementary characteristics of target multimodal features and multi-model iteration have been fully utilized to realize the improvement and optimization of the intelligent classification and detection algorithm for small targets.The intelligent neural network for infrared small target detection has been designed,which could improve the robustness and dynamic adaptability of the detection algorithm,and effectively improve the detection accuracy of small targets.The research contents of this dissertation mainly focus on the following four aspects:Firstly,in the optical imaging characteristics domain,in order to solve the problem of low detection accuracy caused by low imaging pixels and the lack of shape and texture features of small targets,an interframe continuous feature extraction model based on the optical imaging feature of small targets is proposed.An image sequence feature extraction network combining multi-scale convolution feature fusion and attention mechanism is constructed based on the temporal coherence of the continuous frames.The dilated convolutional block attention module integrated with Resblock and the skip connected feature pyramid network are proposed to fully extract the multi-scale feature map of small targets.The above convolutional neural network takes into account the synergetic enhancement of the completeness of semantic information and the fine granularity so as to improve the detection accuracy of small targets in the optical imaging feature extraction and classification network.Secondly,in the radiation and motion characteristics domain,in order to solve the problem of low detection accuracy caused by the low accuracy of the infrared radiation characteristic measurement,a prediction model of atmospheric transmittance based on deep convolution neural network is constructed.The infrared radiation characteristics parameters with low error accuracy are obtained through the radiation characteristics inversion as high-precision radiation characteristics data sources for target detection model to improve the detection accuracy.This approach breaks through the technical bottleneck of difficult points in the prediction of atmospheric transmittance and low accuracy of characteristics measurement in complex environments.The empirical formula of atmospheric transmittance calculation is derived based on the oblique atmospheric transmittance model and the least square method.Furthermore,the deep convolution neural network model on atmospheric transmittance prediction is put forward accordingly in which the features of multi-dimensional infrared measurement sequence for the atmospheric transmittance calculation could be fully extracted.Both of the accuracy of atmospheric measurement and the target detection are increased compared with the traditional atmospheric calculation software.In order to solve the problem of low detection accuracy caused by the poor feature extraction ability of infrared measurement,a feature extraction model based on multiscale convolution neural network for infrared measurement sequences of small targets is proposed in which the infrared radiation characteristics and motion characteristics are used as inputs.In order to extract the complex and scattered multi-dimensional infrared measurement sequence features,firstly,the multi-scale dilated convolution module and convolution neural network are combined to improve the expression ability of the target local spatial sequence features.Secondly,the bidirectional encoder representation from Transformers structure is used to learn the global depth features between different spatial positions of the sequence features.The robust classification ability of the infrared small targets is effectively improved compared with the sequential feature extraction model based on recurrent neural network.Thirdly,in the multimodal domain,in order to solve the problem of low target detection accuracy caused by low signal-to-noise ratio of infrared small targets,lack or even missing of shape texture structure,and insufficient feature description of small targets by single modal data,the optical imaging characteristics,infrared radiation characteristics and motion characteristics of small targets are used as data sources.Firstly,the detection method is optimized according to the quality of input data sources to determine the most efficient target detection model.Then,a hybrid multimodal feature fusion and classification detection network is proposed combining feature fusion and decision fusion.The complementary characteristics of multimodal data and multiple model iterations are made full use of and the multimodal features of targets are deeply extracted.Rich feature vectors of the small targets are obtained and the robustness and dynamic adaptability of the network are improved.Compared with commonly used classification and detection algorithms in engineering,it effectively improves the detection accuracy of infrared small targets.Efficient matching of detection performance and speed are achieved in practical engineering applications.To sum up,the difficulties in the detection technology of small targets are analyzed and summarized and the relevant technologies and theories involved are studied in this dissertation.A framework of the classification and detection of infrared small targets based on neural network model is built.The algorithm has achieved engineering application and performance verification on a certain type of vehicle mounted multiband early warning equipment for the first time,and the average recognition rate is better than 90% which meets the actual operational requirements.It provides a theoretical basis for the follow-up research on infrared small target detection technology,and has broad military and civil application prospects.