Research On Deep Convolutional Neural Network For Small Target Detection In Infrared Image

Because of its good concealment and strong anti-jamming ability,infrared detection technology has been widely used in infrared guidance,early warning and other military fields.Among them,the(weak)small target detection ability in infrared image directly determines the level of modern weapons and equipment,which has become one of the key research and development directions.How to solve the problems that the infrared dim and small targets occupy very few pixels in the infrared image,the features are not obvious,and the target features are easy to be confused by noise,and how to effectively improve the detection ability of infrared(weak)small targets is the current research focus.This paper studies the detection algorithm of air-to-air infrared weak and small targets and ground-to-ground infrared small targets based on depth learning.The main contributions of this paper are as follows:1.This paper briefly introduces the mathematical basis of target detection based on deep learning.A novel infrared dim target detection algorithm,called J-MSF,based on multi-channel and multi-scale feature fusion is proposed,which solves the problem that the classical infrared dim target detection algorithm based on deep learning cannot detect because the target information disappears in the upper receptive field.Firstly,a new multi-channel Janet structure is proposed to design the J-MSF backbone extraction framework.Secondly,a descending threshold feature pyramid pooling structure(DSPP)is exploited,and a multi-scale fusion detection strategy is conducted.Finally,the Gauss loss optimization function is designed.The experimental results show that the recall rate and the AP value of the proposed algorithm are improved by 9.07,9.89 and 1.67,3.16,respectively,compared with those of YOLOv3 and YOLOv4 algorithms in "A dataset for infrared detection and tracking of dim-small aircraft targets underground / air background ".2.A multi-layer fine-grained perceptual network detection algorithm based on self-coding learning(MFG-TF)is proposed to solve the problem that the target is easy to be confused with background noise and lead to low detection accuracy when the target detection algorithm of ground infrared small target based on depth learning is used to detect targets with uneven thermal radiation.First of all,the algorithm uses the self-coding and decoding structure of CNN and Transformer to build the backbone extraction network;secondly,a scale adaptive fusion module with built-in expansive convolution is designed;finally,the Transformer detector is used to detect small targets.The experimental results show that the detection effect of the proposed algorithm MFG-TF on the infrared time-sensitive target detection and tracking data set for air-to-ground applications is 3.85,1.18,1.32 and 2.26,0.66,0.94 respectively compared with YOLOv3,YOLOv4 and Vi T algorithms.In this paper,the algorithm is better than the current mainstream algorithms in infrared(weak)small target detection,which shows good real-time performance and adaptability,and can be effectively applied to air-to-air infrared small targets and ground-to-ground infrared small targets.At the same time,J-MSF and MFG-TF also show good robustness and high detection performance.