Research On Infrared Small Target Detection Using CNN Based On Region Proposal

With the advancement of military modernization,infrared detection equipment has been widely used in equipment night vision,target reconnaissance,missile interception and other fields.The detection and tracking of small and weak targets in infrared images is conducive to expanding the scope of equipments and fighting for favorable combat opportunities.Infrared small target detection has gradually become a key technology in the battlefield.However,the current mainstream infrared small target detection algorithm mainly aims at the sky background,and the false alarm rate of the algorithm is high in the complex background which appears special objects;the common deep learning target detection algorithm mainly aims at the natural image,and is not suitable for detecting infrared small targets directly.After fully summarizing the previous research,this thesis uses traditional algorithms to extract local areas where targets may exist,and applies the convolutional neural network to classify small target candidate regions,and achieves detection of infrared small targets in complex backgrounds with a high recall rate and a low false alarm rate.The main work and contributions of this thesis are as follows:(1)Ten thousand infrared small target images are captured and made into dataset.There is a lack of public data sets in this field,and the amount of existing data is not enough to support network training.In this thesis,a large number of flying infrared small targets in complex background have been photographed in more than one year,and the infrared dataset containing 10420 images,10545 targets and 4 kinds of background has been made.In addition,the methods of clipping and embedding are used to achieve the data enhancement,which effectively balances the number of samples in the target and the background regions before training.(2)The traditional infrared small target detection algorithm is improved to region proposal algorithm.In order to extract local regions which contain tiny and low signal-to-clutter ratio targets,this thesis improves the disadvantage of the original IGVF algorithm: high false alarm rate when adjacent hot-pixels and dead-pixels appear,which reduces the average false alarm rate by 11.5% and increases the average recall rate by 24.3%.Through accelerated by GPU,the improved version speed is 96 times faster than the original algorithm.Finally,the candidate region center is extracted by NMS algorithm,which is also suitable for other traditional algorithms.(3)An infrared small target region proposal algorithm based on the fusion of segmentation and traditional methods.This thesis innovatively introduces the segmentation algorithm into infrared small target detection,uses the area and aspect ratio to obtain the small point areas that may contain small targets,and then merges them with the pixel area extracted by the improved IGVF algorithm.Finally,tested by our dataset,the region proposal algorithm with high recall rate of over 90% is achieved.(4)Convolutional neural classification network for small infrared targets.In order to improve the classification accuracy of small target candidate regions by CNN,this thesis designs a large-scale network entrance,and uses FPN to fuse the context semantic features,outputs the prediction results of the network on the large-scale feature map.Tested by our dataset,the recall rate and accuracy rate of CNN classification are both over 95%.Finally,the target is located by the center of the candidate region,and the low false alarm rate of infrared small target detection in complex background is achieved.