Research And Implementation Of Infrared Small Target Detection Based On Deep Learning

As a branch of the object detection field,infrared target detection has incomparable advantages over ordinary target detection.Due to the use of infrared thermal imaging technology,infrared target detection can work around the clock,with a long view distance and strong anti-interference ability.It is precisely because of these advantages that infrared target detection is widely used in military,industrial and civilian fields.However,the target account for a small proportion in the image when the imaging distance is far.The target?s brightness is low,and often appears as a weak target.Small and dim targets occupy fewer pixels in infrared images,their features are not obvious,and are easily interfered by noises such as clutter and heat sources.This makes it very difficult to extract features based on deep learning methods.As a result,the detection accuracy is low,the location is easy to shift,and the missed detection is serious.In response to the above problems,this paper starts research from the perspective of feature extraction and data balance.The network can perceives the location of small and weak targets for effective learning by increasing the IoU prediction module;we introduce a better loss function to control the impact of the imbalance of positive and negative samples;we increase the number of small and small targets in the picture to promote the extraction of the characteristics of small and small targets by the network,So as to solve the problem of poor detection effect of infrared weak and small targets in complex infrared scenes.The main work and innovation are as follows:(1)The thesis introduces the related knowledge of deep learning algorithm and infrared small target detection and the current research status at home and abroad.The performance,advantages and disadvantages of traditional infrared small target detection,as well as the advantages of deep learning in small object detection are discussed,and the method is applied to infrared images to detect small and weak targets,and how to improve the detection ability of the model is discussed.(2)Aiming at the difficulty problem of small infrared target detection and location,an improved infrared small-scale target detection algorithm based on RefineDet network is proposed.First,we add an IoU prediction module to the network,so that the network can perceive the location of weak and small targets,and then calculate the prediction loss to optimize the training process of the model and perform more effective learning.Then,aiming at the problem of too many negative samples in anchors,we introduce a better loss function to improve the impact of sample imbalance by assigning different weights to different categories.Experimental results show that the improved algorithm hardly affects the detection speed and effectively improves the accuracy of small-scale infrared target detection.(3)Aiming at the problem that weak and small targets are prone to missed detection,a simple and easy-to-implement object moving algorithm is proposed.By moving the small and weak targets in the image,the number of weak and small targets is increased indirectly,so that the network can obtain enough weak and small target features for learning,so as to improve the detection ability of the model.The experimental results show that the object moving algorithm can improve accuracy and reduce missed detection and false alarm.Especially through targeted target removal,the missed detection rate can be effectively reduced.(4)In order to promote the practical application of the research,this article uses mice as the detection targets,and builds an infrared small target detection system,namely the infrared mouse detection system.By outputting the position of the mouse in the scene and the time when it appeared,the habits of the mouse in the scene can be obtained,and the mouse can be accurately eliminated,so as to provide detection methods and technical support for preventing the rat disaster.