Research And Application Of Object Detection Algorithm Based On Deep Learning

With the continuous development of computer vision,target detection is one of the most popular research methods in this field.Whether it is in visible light or infrared scenes,it has a very wide range of applications in many fields,such as military reconnaissance,unmanned driving,intelligent monitoring and so on.At present,the use of deep learning-based target detection algorithms under visible light conditions has achieved rapid development,while most infrared scenes still use traditional methods.This uneven phenomenon makes the development of target detection in infrared scenes very limited.Compared with visible light,some features,such as color and texture,are lacked in infrared images.Generally,infrared target contour is blurred,the contrast is low,and it is also easily interfered by noise,resulting in poor detection results.The deep learning-based methods have powerful feature learning capabilities and high detection accuracy,and in order to be implemented in practical applications,the speed needs to be optimized.Therefore,based on the deep learning detection algorithm YOLOv3,a series of improvement strategies for real-time target detection of ground targets such as vehicles and pedestrians in infrared scenes are proposed.The main contributions of this thesis are:(1)In order to further improve the detection speed of the deep learning algorithm,on the basis of the lightweight network MobileNet,an improved lightweight algorithm basic network LMobileNet is proposed.First,for depth-wise separable convolutions,the parameters and calculations amount of the module are reduced by channel compression and grouping convolution.At the same time,because the information exchange between feature channel will be blocked in the grouping convolution,an improved channel shuffling method IShuffle is introduced.Then,combined with the residual structure to improve the generalization ability of the network,the RLDWS module was constructed in combination.Finally,the RLDWS module will be used to form a more lightweight network LMobileNet.Experiments show that the classification accuracy of the improved network is improved by 22.5%,while the amount of parameters and calculations are only 31.7% and 52.6% of the original.It also has certain advantages compared with other lightweight networks.(2)For the problem of miss detection and inaccurate positioning of multi-scaletargets in the infrared scene of YOLOv3 algorithm,an improved infrared target detection algorithm LYOLOv3 is proposed.Firstly,it is combined with the improved algorithm basic network in(1)to replace the original darknet-53 basic network,the algorithm detection speed is improved.Then,by introducing an improved feature fusion method IFPN to increase the application of shallow features,the detection effect of small targets is improved.And the DGIOU loss function based on the GIOU loss function is improved to improve the problem of inaccurate positioning.Experiments show that the improved algorithm improves the detection accuracy by 11.3%,and can maintain the speed of 52.5 frames per second,which effectively improves the detection performance in infrared scenes.(3)Based on the NVIDIA Jetson TX2 platform,the improved LYOLOv3 algorithm is implemented and optimized.The infrared data set was constructed first,and then the automatic test software was designed according to the system requirements.Then,the algorithm was optimized on the TX2 platform.Finally,the practical value of the system was demonstrated through the test.