Infrared Small Target Detection And Embedded Deployment Based On Deep Learning

Different from radar and visible light imaging,infrared imaging system belongs to passive imaging,mainly using the characteristics of infrared radiation in all objects in nature,so it has strong penetration,night imaging and recognition camouflage,all-weather work,long distance and other advantages,so it has been widely used in military and civilian fields.In terms of military defense and guidance,the target detection system can accurately and timely identify long-range air strike targets as a guarantee for follow-up response work.Therefore,the detection of small targets in infrared images has become an extremely important study in the early warning system.Only timely and accurate discovery of the target location can ensure the smooth development of subsequent identification,tracking and other work.With the increase of the distance between the target and the detector,the infrared small target presents the characteristics of weak target in the image,and lacks detailed information such as texture and shape,and the background is complex and changeable,resulting in the difficulty of detecting the small target submerged in it,and the bright spot of the target generated by the blind element in the background is easy to lead to problems such as algorithm misdetection.Although a large number of researchers have studied these issues,there are still some problems that are not well addressed.(1)Aiming at the problem of small infrared target size,complex and changeable background,and serious interference of target bright spots generated by blind elements,we propose infrared small target detection based on global feature association learning.The algorithm consists of multiple modules,the core part of which is to use the global feature association method to eliminate the influence of the most difficult target bright spots in infrared small targets,design spatial and channel attention to strengthen the network’s attention to the characteristics of small targets,and avoid feature misalignment during small target feature fusion through the anti-aliasing context feature fusion module.Under the same experimental conditions,the proposed method is compared with other methods,and the detection results of each method are visualized,which verifies the effectiveness of the proposed algorithm.(2)For the problem that the large number of deep learning models,the large amount of computation and the large number of model parameters are not conducive to the deployment of embedded devices,we design a lightweight deep learning infrared small target detection model Lite Seg for deployment,which is mainly constructed by deep separable convolution,the model parameters and floating-point arithmetic amount are greatly reduced compared with traditional neural networks,and the inference speed of the model is also significantly improved.The forward inference framework Tensor RT was used to build the inference engine,and the proposed model was successfully deployed to the embedded device Jetson Tx2.