Infrared Dim Small Target Detection Based On Joint Temporal-Spatial-Spectral Features

Infrared dim and small target detection is the key technique in the infrared surveillance system,infrared early warning system,infrared search and tracking system and so on,and is widely applied in the fields of traffic,security and military.Due to the long imaging distance and the interference brought by noise and background clutters,in these applications,the targets in infrared images appear dim and small.The infrared dim and small target has the following characteristics: lack of features,low brightness,complex imaging conditions,complicated motion pattern and so on,and these factors bring challenges to the detection task.After decades of development,infrared dim and small target detection technique has made great progress,but in the complex background conditions or diverse infrared scenes,the existing detection methods are difficult to remove false alarms stably,and still need to be improved in the practical applications.In this dissertation,we research on the infrared dim small target detection method based on joint temporal-spatial-spectral features.We devote to searching and constructing features in the spatial,temporal and spectral domains so as to improve the separability between targets and backgrounds,and utilize the multi domain features synthetically in order to improve the quality and stability of the infrared dim small target detection.In concrete terms,the thesis mainly includes the following contents:(1)We introduce the theoretical basis of the infrared dim and small target detection.We analyze the characteristics of infrared targets and backgrounds,and give the description,the framework and the evaluation metrics of the infrared dim small target detection task,in comparison with the common object detection task.These research contents are applied in the other parts of the thesis.(2)We propose an infrared background clutter suppression method based on chaingrowth filtering.The chain-growth filtering model can adjust its shape to adapt to various clutter structure,such as lines,curves,irregular edges and so on,and can only involve the pixels inside the clutters in computing,weakening the influence of clutter shape on feature extraction.Compared with the feature extraction strategy based on fixed shapes,the chain-growth filtering model has a stronger ability to suppress the infrared background clutters,so it has a better detection performance.(3)We propose an infrared small target detection method based on density peaks searching.We map the pixels into the “density-? distance” feature space,eliminate the pixels in flat background regions fast by density peaks searching,and extract the points that have local maximum density(density peaks)as candidate targets.For these candidate targets,a maximum-gray region growing method is proposed to suppress the background clutters with complex shapes.In the decision stage,a quartile-based technique is adopted to obtain a more robust decision threshold.In the process of performance evaluation,the density peaks in infrared images are divided into positive and negative samples,and PR curve can be drawn by changing the decision threshold;compared with the ROC curve that commonly used in the field of infrared dim and small target detection,PR curve can reflect the performance differences between the algorithms more clearly.This proposed method maintains the high quality and efficiency of feature extraction,and does well in both detection performance and operation speed.(4)We propose an infrared dim and small moving target detection method based on kernel correlation filtering.The background motion is modeled and compensated by kernel correlation filtering,the relative motion feature of current frame is extracted by frame differencing,and the detection results can be obtained by thresholding the relative motion feature map.This method achieves the separation of background and target motion modes.The proposed method effectively utilizes the relative motion features of the targets and greatly improves the detection performance under the moving background condition.In addition,the method puts most calculations in the frequency domain,and achieves a high efficiency that could meet the real-time requirements,which benefits the practical application.(5)We propose a robust detection strategy based on the combination of features in spatial,temporal,and spectral domains.This method uses chain-growth filter to extract spatial features,uses kernel correlation filter(mainly in frequency domain)to extract temporal features,and utilizes these two independent and complementary features to jointly express each pixel in infrared image.In the extracted multi-domain features space,anomaly detection algorithms are used to calculate the confidence of each pixel.The final detection results are obtained by thresholding the confidence map.Taking advantages of the features in spatial,temporal and spectral domain,the proposed method increases the separability between the targets and backgrounds in the feature space,improves the confidence of detection results,and strengthens the robustness of detection in various types of infrared scenes.According to the results of extensive experimental tests,the spatial detection methods and temporal detection methods proposed in this thesis perform well in comparison with other same type methods,improving the performance of the detection under complex background conditions;the multi-domain-feature-combined detection method proposed in this thesis performs well in comparison with other single domain algorithms,improving the stability of the detection in different types of scenes.In addition,the methods proposed in this thesis think highly of the efficiency,which makes the study have both theoretical significance and application prospects.