Object Detection In Infrared Scenes Based On Deep Learning

Infrared imaging systems are widely used in many fields such as intelligent driving,infrared reconnaissance,and environmental monitoring due to their all-day characteristics.With the improvement of GPU computing power and the emergence of large-scale labeled datasets,it is of great significance to use deep learning methods to research infrared target detection.However,due to the long wavelength of infrared images,the outline of the target is blurred,the feature details are not obvious,and the color information is lacking,which is more difficult to detect compared with visiblelight images.In fact,although infrared target detection has made great progress with the rapid development of deep learning networks,there are still the following problems:(1)The existing convolutional activation structures lack flexibility when extracting features,and the encoding capacity is limited by the(2)The convolutional neural network mainly models local pixels and does not consider the feature relationship between global pixels;(3)The saliency of the infrared target relative to the background is not fully utilized,and the location information attention is lacking;(4)When the embedded platform is deployed,the network reasoning speed is slow and cannot achieve real-time detection.Given this,this paper carried out the following research work:(1)In view of the lack of infrared target feature information and the poor feature extraction ability of ordinary convolutional network structures,this paper designs the Effi-YOLO network based on the YOLOv3.The network uses a dynamic convolution activation structure to flexibly adjust the convolution calculation area and weight coefficients according to the input features and adopts an adaptive activation rate according to the calculation results.A lightweight and efficient backbone network is used for basic feature extraction,and a significantly enhanced receptive field module is used to expand the model receptive field,establish a new target position loss function,and improve the model target regression positioning accuracy.In the FLIR dataset test,the size of the new model is reduced by 33.3% compared to the baseline algorithm YOLOv3,and the detection m AP is improved by 9.9%.(2)The DINet network model is designed to solve the problems that the convolutional neural network only models local pixels,lacks long-range dependency information between pixels,and does not fully utilize saliency information.DINet combines the global information of the Transformer and the local information of the convolutional neural network to strengthen feature extraction and modeling capabilities.The image input stage utilizes the saliency prediction network to generate pseudo-color images with saliency target information.A new receptive field enhancement module is constructed to enhance the saliency information inside the target area while expanding the receptive field of the model,and further improving the infrared target detection performance through the multi-layer feature layer fusion structure.In the FLIR dataset,the detection accuracy is 5.5% higher than that of YOLOv5-S;In the KIAST dataset,the missed detection rate is 4.11% lower than that of IATDNN+IASS.(3)In order to meet the needs of practical engineering projects,an infrared UAV dataset is constructed for the training and verification of the network model.Choose the lightweight YOLOv5-S network and transplant it to the TX2 platform after training on the desktop.To further accelerate the model inference speed,Tensor RT acceleration technology is used on TX2 to optimize floating-point calculation and model structure fusion on the transplanted YOLOv5-S,which improves the model inference speed multiple times.The final inference speed reaches 58 frames per second,meeting the real-time detection requirements.In summary,the research has effectively improved the target detection capability and the target detection performance of the infrared system through efficient feature extraction,adaptive dynamic convolution activation structure,and feature enhancement of dual-modal images.The verification results show that the algorithm is significantly improved compared with the baseline algorithm.