Dim And Small Target Detection Based On Deep Learning

Dim and small target detection is closely related to national security.It is widely used in the fields of airspace security,airport clearance protection and air control in military sensitive areas.The small target has small imaging area,weak feature information,and a wide range of scale changes.These characteristics make it difficult to be accurately detected in the real-time system.Therefore,how to capture targets in real time and accurately from a complex background then issue timely warning is the key of this technology,which is also the focus of the research in this topic.Based on this issue,this thesis did the following:Based on the traditional dim and small target detection method,this thesis proposes a multi-scale filter fusion detection algorithm to effectively solves the detection problem that the traditional filter algorithm is difficult to be applied to the target scale variation.Firstly,constructs an image pyramid by downsampling the image at different scales,then uses a cascading filter for background suppression in each scale space and adopts an adaptive threshold segmentation method to obtain the target information,and finally fuses the target information in each scale space to achieve multi-scale target detection.This method effectively reduces the calculation amount of the algorithm by combining the detection method of cascade filter and image pyramid.Traditional dim and small target detection methods are mostly based on artificial rules,and in practice it is difficult to find global matching rules,so this type of method often only performs well in a particular scenario,but is difficult to apply to other scenarios.The target detection algorithm based on deep learning allows the model to autonomously select features during the training process and continuously optimize the selection through data-driven modeling,so that the model can learn features in various scenarios.This thesis is based on the lightweight target detection network Mobile Net V2＿SSD,improves the ability of model to express small target features by adding the feature pyramid structure.The experimental results show that the improved model has similar detection speed as the original model,but the detection result on small targets is significantly improved.The Autoencoder can achieve accurate reproduction of the input signal.Based on the characteristics of small target images,this thesis proposes to use a convolutional autoencoder network to reconstruct the background of small target images,train the network to map the input image into a background image without the target,and then obtain target information through image difference and image segmentation techniques to achieve real-time and accurate detection of small targets.In this thesis,a convolutional layer with a step size of 2 is used instead of the pooling layer,thereby ensuring the feature extraction performance of the network.Experimental results show that the dim and small target detection algorithm based on convolutional autoencoder has better detection performance than the traditional background suppression filter algorithm.