Research On Deep Learning Method For Infrared Dim And Small Target Detection

Infrared dim and small target detection technology has a wide range of applications in perimeter security,target tracking,fire prevention and disaster prevention and other fields.Infrared dim and small targets are extremely difficult to detect due to their small size,inconspicuous features,and often being interfered by background clutter.Existing target detection methods have serious missed detection and false alarms.In response to these problems,this project carried out research on infrared dim and small target detection methods based on deep learning theory.The main work is as follows:（1）A single frame dim and small target detection method based on Twin Transformer in infrared technology is proposed.In this method,the original image is divided into several non-overlapping sub-images,and then the target sub-image and a set of trained templates containing location information are sent to the Siamese feature extraction network to extract features.Finally,the extracted subgraph features and template features are sent to the feature matching module,and the matching degree of the two is used to determine whether the target subgraph contains the target and calculate the coordinates.This method takes the whole target as the object for feature extraction,this feature extraction method fully considers the context information of the target,making the feature extraction results more sufficient.The experimental results show that the single frame detection method has a higher detection rate in some simple scenes such as prominent objects.（2）A multi-frame fusion detection method for dim and small targets in infrared technology is proposed.The use of multiple continuous data frames,through the multi-frame target fusion module,analyzes the deep semantic features of the background and moving targets in continuous images to identify the moving target,thereby highlighting the target and suppressing the background,and sends the processed feature map to the optimized single-frame detection network to predict objects.The experimental results show that compared to single-frame object detection method,multi-frame target fusion detection method make the average detection rate of targets greatly increase and reduce false alarm rates even in complex backgrounds,demonstrating better detection performance.（3）Sufficient ablation experiments and comparative experiments have been carried out.The project analyzes the possible impact of different parameters in the main modules of the multi-frame fusion detection network on network performance,and different ablation experiments were performed to compare and test in order to achieve the optimal final evaluation indicators of the network,and provide experimental references for the design of similar network tasks.The project selects several existing detection methods to test on public datasets.The experimental results show that the average detection rate and false alarm rate of the detection method proposed in the project are 0.985 and1.96×10^-4,respectively,which improves the average detection rate by28.76%and reduces the false alarm rate by 79.37%compared to 0.765 and 9.5×10^-4 of the best existing method.