Research And Implementation Of Motion Deblur Algorithm For Object Detection

The development of object detection technology has provided convenience for people and greatly improved the performance efficiency of various corresponding tasks.However,due to the non-uniform relative motion between the potential object and the camera lens,the image will be unevenly blurred.The existence of these ambiguities will greatly impress the operation effect of the object detection framework,so it is of great significance to the research of the object detection-oriented motion deblurring framework.This thesis mainly focuses on the motion blur of the image,with the goal of improving the object detection effect of the blur image as the final goal,and researches the image deblurring framework.The specific research contents are as follows:(1)three-channel multi-scale deblurring frameworkIn this thesis,a three-channel multi-scale deblurring framework is proposed to solve the problem of large difference in blur degree in images.By setting up different channels,using different down sampling mechanism and deblur module,the data is discarded to different degrees and reconstructed with different strength.In this way,each channel can better adapt to the corresponding degree of blur,so that each degree of blur in the image can be effectively removed.(2)inter-layer attention mechanismIn order to solve the problem that extra noise may be introduced in the transfer of inter-layer results,this thesis proposes an inter-layer attention mechanism.In the process of inter-layer transmission,since the channel with high repair intensity uses the same operation for the image as a whole,there will be excessive repair when the image block with relatively low blur degree is repaired.If the result of excessive repair content is passed into the subsequent network,the repair effect of the subsequent framework will be reduced due to the introduction of extra noise.To solve this problem,this thesis introduces the inter-layer attention mechanism in the three-channel multi-scale restoration framework,and proposes the attention-based multi-scale content deblurring network(Attention-based Multi-scale Network,AMNet).(3)two-stage deblur strategy based on contour generationThe Edge Generating-based two-stage Deblurring Framework(EGTDFram)is proposed in this thesis because the object detection needs the input with a clear edge.Because the deblur involved in this thesis will contain a large degree of deblur,so the noise and signal mixing are serious,it is difficult to directly obtain the potential clear contour from the blur contour extracted from blur image.In order to solve this problem,this thesis directly extracted the information needed for the potential clear contour from the original blur image and reconstructed the contour,and then combined with the content deblur network to generate the potential clear contour.(4)blur image object detection systemThis thesis designs and implements the blur image detection system,which includes the functions of image deblurring,blur image detection and deblur image detection.By embedding the framework of this thesis into the application system,the accuracy of blur image detection is improved,and the practicability of this framework is verified.