Infrared Dim Small Target Detection Based On Deep Learning

The detection of infrared small targets is a key technology in the field of infrared guidance.It plays an important role in the aerospace research field such as infrared stealth spacecraft detection,small celestial body detection,missile guidance,and battlefield reconnaissance.Taking the technical requirements of remote detection of infrared small targets into consideration,the target detection and trajectory tracking of infrared images under complex backgrounds is researched here.The main work and research results of this thesis include:Firstly,the application background and significance of infrared small target detection technology in the detection and defense of small celestial bodies in deep space exploration missions is recalled.In addition,the research contents of moving small target detection from ground-based platforms and space-based platforms were described respectively,followed by an introduction of the current research status of detection technology.Then,the model construction methods of infrared image is described.According to a detailed analysis of the target's radiation characteristics,background characteristics and noise characteristics,the performance evaluation indicators of small target detection is presented and the difficulties of infrared small target detection is summarized.Besides,three detection algorithms based on human visual characteristics are introduced as comparative methods to verify the scientificity and effectiveness of the proposed algorithm.For the targets with SCR less than 3 in a single infrared image under complex environment,traditional algorithms with manual feature extraction is prone to false alarms,while deep learning with powerful feature extraction capabilities cannot train tiny targets with lack contour information.In this context,sliding window sampling training method is adopted,which originates from the idea of nested structures in traditional algorithms based on human visual characteristics,and a fully convolutional network using recursive convolutional layers is designed to extend the depth of the network without increasing training parameters.The multi-branch structure of the network's parallel convolution structure simulates the multi-scale operation of the traditional algorithm.Additionally,various loss functions are designed to combat the serious imbalance of positive and negative samples.The results show that the algorithm achieves better detection performance than traditional algorithms.For the weak targets with less than 12 pixels in a single image,the chain structure will lose the target information when the network is deepened.However,the encoderdecoder with information fusion and supervision mechanism can significantly reduce this defect.On this basis,a fully recursive convolutional network is designed.By drawing on the characteristics of encoder-decoder feature fusion,the proposed algorithm take the benefit of sliding window sampling training in fully convolutional network and uses dense connection operations,recursive convolution operations and parallel convolutions operations.It is shown that the algorithm achieves better detection results than traditional method and the fully convolutional network in this paper,which mean target detection rate of this method is always the highest under the same false alarm rate.Finally,for the moving target with less than 6 pixels in the image sequence,a model based on 3D convolution kernel and convolutional long short-term memory layer is proposed.where the fully-connected operation in the gate unit of the long short-term memory layer is modified by convolution operation.Aiming at the problem that the model has a lot of residual noise,an attention mechanism is brought in.The 3D convolution kernel extracts the short-term spatio-temporal information of 15 consecutive images,the convolutional long short-term memory layer extracts the longterm spatio-temporal information of the sequence,and the attention mechanism discards the background information and pay attention to the target information.Compared with other methods,the simulation results show that the convolutional long short-term memory network based on attention mechanism with output gate has an average reduction of 31.0% and 39.5% in root mean square error and average absolute error,and an average increase of 18.7% and 3.1% in peak signal-to-noise and structural similarity.So it proves that the algorithm can detect the trajectory of infrared target with less than 6 pixels,while the clutter of background is the least and the detection result is the most excellent.