Research On Target Detection Algorithm Of Through-Wall Radar Based On Deep Learning

Ultra-Wideband Through-Wall Radar is a new type of all-weather,all-day radar imaging system in close range.It can not only easily detect targets through walls and ruins,but also help users obtain useful target information.Therefore,it is widely used in special occasions such as post disaster rescue,anti-terrorism capture,military operations and so on.However,Through-Wall Radar image based on the measured data usually has problems such as multipath interference and grating lobe,which hinders the detection of targets.Based on this background,the method of deep learning will be used to conduct in-depth research on the two key technologies of clutter suppression and target detection and location.Firstly,aiming at the problem of Through-Wall Radar image target detection,this thesis proposes SE-DTYO5 Net based on channel attention mechanism and multi-scale feature fusion.The model is based on YOLOv5 framework and embedded with parallel computing DETR module,which improves the computational efficiency of the model.By introducing the idea of SENet feature channel attention,the ability of beneficial feature learning and the pertinence of feature learning are enhanced.At the same time,by adding YOLO branches and adopting multi-scale feature fusion strategy,the model can not only improve the flexibility of the receptive field,but also strengthen the re-integration between features and improve the learning ability of abstract features and location information.The experimental results show that the model performs well on the TWRDS data set,and its AP₅₀ value can reach 87%,which meets the expected goal of the model improvement.Secondly,since the radar image generated by the actual measured radar echo data after BP radar imaging technology still contains a large number of signal clutter,such as multipath interference and grating lobe,which seriously affects the effect of target detection.To solve this problem,a Swin-SK Block module based on optional convolution and shift window mechanism is proposed.And taking this module as the core of feature extraction and reconstruction,a clutter suppression algorithm for Through-Wall Radar image UST-Net is designed.This algorithm uses the strong learning ability of the attention mechanism and the neural network,takes the clean Through-Wall Radar image as the reference,takes the real Through-Wall Radar image with multipath interference and clutter as the model input,and continuously experimental trains,so that the final model outputs clean radar image.By designing a comparative experiment,the thesis verifies the clutter suppression ability of the model on the Through-Wall Radar image carrying clutter,and its evaluation index value is about 23%higher than that of the denoising autoencoder.Finally,this thesis connects UST-Net and SE-DTYO5 Net in series.The Through-Wall Radar image with clutter is sent to UST-Net for clutter suppression,and clean through-wall radar image is obtained.Then,it is input into SE-DTYO5 Net network for target detection.Experiments show that the target detection effect of Through-Wall Radar image processed by UST-Net is much better than that of unprocessed image.