Study On Infrared Target Detection Technology Under Weak Feature Conditions

Infrared imaging systems have the advantages of anti-interference,all-weather imaging,high concealment capability,and accurate target detection at long distances,which make them very widely used in civil,military,and industrial fields.However,compared with visible images,infrared images have the disadvantages of low resolution,blurred edge and high noise,which can lead to low accuracy and poor robustness of target detection algorithms for infrared images.In this thesis,by analyzing the performance of mainstream visible light-based target detection algorithms under infrared data sets,designing and improving the network structure,data processing methods,and training strategies,we propose a detection algorithm for weak targets in infrared images,while pruning and accelerating the model,and realize the deployment on NVIDIA Jetson Xavier embedded platform to complete the high-precision real-time detection of weak targets in infrared.The details of which are as follows:1.Adopt high bit width infrared image to detect and process.Explore the principle of infrared imaging and infrared characteristics,compare with visible images and summarize the advantages of high bit width and the shortcomings of low resolution and low contrast of infrared images.Surf modeling and simulation of image data are performed to explore the effect of dynamic range compression methods on weak features in images.By recording ablation experiments,the average accuracy of the Bicycle category increases by 4.8% when using 14 bit raw IR data for training and detection,and the detection accuracy of small targets is significantly improved.2.Design of pre-processing methods for high bit-width infrared images.The applicability of enhancement and augmentation methods based on visible images to infrared images is analyzed by PSNR(Peak signal-to-noise ratio),Average Gradient and other indexes.,To solve the problem of weak feature information in infrared images,the data enhancement and augmentation methods are proposed for 14-bit raw infrared images to accelerate the convergence rate of the network during training.The improvement of network performance by different data enhancement methods is analyzed by validation and testing on FLIR extended dataset.3.Redesign and improve the deep learning network for infrared image features.To address the problem of difficult infrared image feature extraction,the feature extraction capabilities of ConvNeXt,Darknet and Vision Transformer were compared separately,and three different attention mechanisms were added while improving the structure of the ConvNeXt network.The improved network is trained with the prevalent target detection network on the FLIR dataset and self-built dataset,84.9% m AP is obtained with guaranteed Parameters and FLOPs/B(Floating point operations)of 37.3M and55.4B,which is a 3% m AP(Mean average precision)improvement with respect to the YOLOV5 m network.4.Complete the compression and landing of the algorithm.To realize the practical engineering application of the algorithm,we prune and compress the algorithm to reduce the network size and graphics memory overhead without affecting the detection accuracy.Evaluate the flexibility,reliability and cost control of different deployment schemes,establish the target platform,and complete the migration,deployment and practical application of the algorithm.