| Object detection technology is widely used in many fields such as intelligent driving and intelligent security,and improving detection accuracy has always been the goal pursued by researchers.In practical applications,the occurrence of weather such as rain,snow,and haze can have a negative impact on object detection,leading to a decrease in detection accuracy.Existing object detection networks can also experience domain drift under complex weather conditions.To address these issues,this paper proposes a weather image restoration algorithm and a domain adaptive object detection algorithm,and designs a complex weather object detection system based on them.The main research contents of this paper are as follows:(1)To address the problems of existing image restoration algorithms,such as the need to model each type of weather under complex weather conditions and the cumbersome process,a unified image restoration network based on an encoder-decoder structure was designed.A detail extraction module was added to the encoder for restoration of small-scale degradation,and a spatial reduction module was added to reduce computational complexity.A weather type query was introduced in the decoder to obtain task-specific feature vectors,which were then restored into a clean image using the deconvolution operation.Experimental results demonstrated the effectiveness of the proposed network.(2)To address the issue of performance degradation of existing object detection models under complex weather conditions,a domain adaptive object detection network with a three-branch structure was proposed.First,data augmentation was applied to the source domain data.Then,a method of sample pair matching and filtering was proposed.Finally,a domain adaptive network based on distillation learning was designed to train the object detection model.Experimental results on different datasets demonstrated good detection accuracy of the proposed network.(3)A complex weather object detection system was implemented,which includes core functions such as image restoration and object detection under different weather conditions,as well as basic registration and data logging functions.The effectiveness of the system was validated through testing. |
