Research On Object Detection In Low-Light Image Based On Iterative Intensity Enhancement Network

With the rapid development of deep learning technology,object detection technology is increasingly mature and widely used in real life.However,object detection in low-light image is still a very challenging problem.There are many imperfect imaging factors in low-light level image,such as low brightness and serious noise,which hinder the target detection model to learn stable and differentiable features.In order to further improve the detection accuracy of low-light level image,this thesis starts from the two aspects of low-light level image enhancement technology and target detection technology,in order to improve the effect of target detection on the current low-light level image.The main research contents of this thesis are as follows:(1)A low-light image enhancement algorithm based on parameter estimation is designed to improve the quality of low-light image.Iterative Intensity Enhancement Network(IIEN)is designed to solve the problems existing in previous low-level image enhancement methods,such as noise,color distortion and large number of model parameters.In order to solve the problem of noise interference,this thesis designs a pixel level component estimation branch,which can improve brightness while alleviating noise interference by iterating component processing.In order to preserve the image information and reduce the number of model parameters,a resolution-preserving enhanced network structure is proposed,and a specific network is constructed by means of separable convolution and reciprocal residual structure.In order to optimize the final image,this thesis uses the zoom point attention to build the overall optimization branch,which is used to estimate the parameters of color correction matrix and gamma correction,and uses both to improve the imaging effect.It can be seen from the visual effects of experimental comparison images and the objective evaluation index values that IIEN network effectively enhances the imaging quality of low-light level images.(2)Through joint training to improve the enhancement algorithm,and design the effective fusion of the original image and enhanced image information algorithm,further improve the accuracy of target detection.In order to solve the problem of inconsistent targets between image enhancement task and target detection task,this thesis conducted joint training after connecting the detection network of low-light level image enhancement network IIEN and YOLOv5 to improve the detection effect,and generated enhanced images more conducive to target detection based on the weights obtained through joint training.In order to combine the information of the enhanced image and the original image to improve the detection accuracy,the fusion algorithm of the input terminal and the middle segment of the target detection network is designed in this thesis.For the problems of insufficient fusion features and lack of adaptability of fusion mode in the input end of target detection,this thesis designs an effective fusion algorithm based on Dense Block module and improves AFF feature fusion and CA attention mechanism.For the fusion in the middle segment of the target detection network,it is necessary to retain more complete features of the original feature information.In this thesis,an effective fusion algorithm is designed based on the channel attention mechanism and residual mechanism.In this thesis,the target detection algorithm based on the image enhancement method can effectively improve the precision of target detection on the low-light level image,and the m AP value increases from 74.14% of YOLOv5 to 78.05%.