| Infrared dim small target detection plays an important role in national defense security and people’s livelihood applications.Due to the target imaging mode,range and imaging background,the infrared source image has detection problems such as lack of detailed information such as color and texture,low target saliency,poor global contrast,and serious background noise and clutter interference.The detection results of existing baseline methods are prone to false alarm and missed detection.In view of the above problems,this thesis proposes three infrared weak and small target detection methods based on information association to solve the problem of weak and small target location,segmentation and weak and small target detection in complex background.The specific work is as follows.(1)This thesis proposes a method of infrared weak and small target detection based on morphological multi-feature information association.Aiming at the problem of locating weak and small targets with low global contrast in the target neighborhood and high intensity false alarm edge interference in the complex infrared background,this thesis constructs a detection mathematical model based on the prior knowledge of high local gray contrast in the infrared weak and small target neighborhood,two-dimensional Gaussian distribution of gray distribution and low similarity between the target and the neighborhood.This thesis firstly designs a local gray contrast measure to extract the target area and suppress the partially flat background area,then constructs a covariance saliency map and a similarity saliency map according to the target gray distribution and the target neighborhood similarity feature,and finally associates the two maps and uses the adaptive threshold segmentation method to locate the weak and small targets.Compared with the baseline method,this method associates multiple morphological feature information for target and background area classification.A large number of experiments have proved that this method has better performance of infrared weak and small target location in complex background.(2)This thesis proposes a method of infrared weak and small target detection based on multi-scale context information association.The network firstly constructs a multi-scale object enhancement module to transform the hand-designed features into a network structure to enhance the real objects,and extracts the shallow details and deep semantic features of the image.Secondly,the network introduces global target response module,channel attention module and multi-layer feature fusion module to aggregate information,select data and decode targets by associating context information and channel information.Finally,the network constructs a multi-loss joint constraint module to combine multiple losses to effectively constrain the network output to solve the problem of infrared small and weak target segmentation under complex background.Compared with the baseline method,this method has higher values of Io U and F measure metrics,and has better ability to segment the contour of weak and small targets and remove false alarms and missed detection.This method has better detection performance and can be explained to some extent.This method is more robust and reliable.(3)This thesis proposes a method of infrared weak and small target detection based on temporal and spatial information association.The method firstly proposes a suspected target extraction network to pre-extract infrared weak and small targets,and uses the loss function constraint to obtain low false alarm output and low missing detection output,and then fuses them.Then the method uses the time domain inter-frame target association network and the space domain target energy accumulation network to enhance the continuous frame target and suppress the background.Finally,the method associates the output information of the spatiotemporal network to get the final detection result,and realizes the detection of infrared weak and small targets in high noise background.Compared with the baseline method,this method combines space-time information synchronization to perform temporal motion information association and spatial energy accumulation enhancement.A large number of comparative experiments show that this method has better detection effect for infrared weak and small target detection in high noise background.Robust experiments show that this method is also suitable for multi-frame weak and small target detection in complex background.This method is more general and reliable. |
