Research On Infrared Object Detection Algorithm Based On Deep Learning

Object detection has always been one of the most basic and challenging problems in the field of image processing and computer vision.Whether in visible light scenes or infrared scenes,object detection has widespread applications,such as autonomous driving,intelligent surveillance and military reconnaissance.At present,the use of deep learning-based object detection algorithms in visible light scenes has achieved rapid development,while the detection of infrared scenes still develop slowly.Compared with visible light images,infrared images lack features such as color and texture,and have limitations such as low signal-to-noise ratio,low contrast,and low resolution.The hand-crafted features used by traditional methods very depends on expert experience and has poor adaptability,while the deep learning-based methods have powerful feature learning capacity,which can improve the performance of detection.Therefore,this paper conducts in-depth research on object detection algorithms in infrared scenes based on deep learning,two infrared object detection algorithms based on deep learning are proposed,and the effects of the two algorithms are deeply analyzed.Considering that the Faster R-CNN model has high accuracy but is time-consuming,the MobileNet network can accelerate the detection process but degrade the performance.We propose an improved RF-Faster-RCNN algorithm which is both efficient and well-performed.Specifically,for the depth separable convolution,firstly,the computation and number of parameters are reduced by channel pruning and grouped convolution.Secondly,the communication of cross-channel information is enhanced by the use of proposed FShuffle in the network.Thirdly,the generalization ability of the network is improved by introducing the residual structure.Finally,the improved deep separable convolution structure is used to build the RFMobileNet basic network and further build the RF-Faster R-CNN infrared object detection framework.The results show that the FR-Faster R-CNN algorithm in this chapter has faster processing speed compared with the R-CNN and Faster R-CNN algorithms.And the detection accuracy has been improved to a certain extent.For the false positives and bounding box deformation of the SSD object detection algorithm in the infrared scene,an improved RF-SSD algorithm is proposed.The backbone network uses the improved basic network RFMobileNet in Chapter 3,which speeds the processing of the algorithm.For the problem of insufficient feature information of infrared objects,the receptive field module(RFB-w and RFB-d)is proposed.Through dilated convolution and multi-scale feature extraction and fusion,more semantic information can be extracted and the overall performance of small objects can be improved.For the problem of bounding box deformation,the SIOU loss function is proposed.Through the calculation of the similarity of the bounding box aspect ratio,the measurement of the bounding box deformation is increased to ensure the consistency of the scale of bounding boxes.The results show that the RFSSD algorithm in this chapter has improved the false positives and bounding box deformation effectively compared with SSD algorithms,and the detection accuracy for small targets has been improved a lot.