Study On High-speed Motion Deblurring And Object Recognition Based On Deep Learning With Memory

Human visual is the main means of information acquiring,and the machine vision simulates the human visual to perceive and recognize the interest objects.With the visual perception which based on deep learning in the stage of intelligent research,the vision recognition simulates the human intelligence is a meaningful research direction.Learning with memory is the human higher-order cognition,and this dissertation will study the visual perception model with memory for completing the motion detection and recognition.Dynamic image recognition has always been a difficult problem in object detection,and this dissertation will focus on the motion images recognition,most of the images are acquired from a high-speed motion camera.The research shows that only the clear images can make the success of object recognition with deep learning method,thus the motion image deblurring is the key to object detection.Firstly,the principle of image generation and the cause of motion blur are analyzed,and a high-speed motion imaging model is constructed.Then the deep learning method with memory is studied to recognize image blur types which caused by complex environment,and the high-speed motion blur image is deblurred.Finally,a deep learning model with memory is studied for object recognition.Specific research work is carried out in the following aspects.The mechanism of image blurring in high-speed motion is analyzed,and the motion imaging model and the simulation system of image sensor are constructed.In the process of motion imaging especially the fast motion,the relative motion between the image sensor and the measured object during exposure time will result in image blur.Therefore,it is very necessary to analyze the cause of the image blur before image deblurring.Through the imaging trajectory analysis of the object motion and the camera motion,the direction of the camera motion is decomposed into parallel to the image plane and along the optical axis of the camera,then the motion imaging model and the motion imaging simulation system are constructed.A deep convolution network model with memory is proposed to recognize the types and parameters of the image blur.The CNN can extract the spatial feature information,but it has the problem of gradient disappearance.The LSTM can solve this problem in the training process of recurrent networks,and it can establish the context dependence of the interest region which is the temporal information.Thus,the spatial information of DCNN and the temporal information of LSTM are fused for image classification and recognition.The experiments show that the fused method is more accurate than using only one of the networks for blur type recognition.By simulating the function of human learning with memory,a recognize model is constructed which based on DCNN fused LSTM.An image deblurring method is proposed especially for seriously motion blur which based on the column grayscale probability consistency.Firstly,the accurate blurred kernel is obtained by the fast Fourier transform twice and the signal accumulation transform.Then the deconvolution operation is carried out for image restoration.Due to the ill-posed problem of deconvolution,the image deblurring will be failed by direct deconvolution.According to the probability consistency of the adjacent columns gray of the natural images,this paper introduces the ? confidence object optimization to optimize the image deconvolution.Finally,a method of predicting boundary blocks is proposed for removing the ringing phenomenon which caused by image boundary information.The experiments show that this deconvolution method is better than the existing methods in dealing with the high-speed motion deblurring.An optimizing method is proposed by using the BRISQUE and CGPC to evaluate and optimize the deblurred image.The existing quality evaluation index of the image restoration is mostly based on the reference image,but the real blurred image has no original reference image.Therefore,it will optimize the restored image which based on the no-reference evaluation index,until the quality index of image restoration is optimized,then stop iteration to obtain the optimal restored image.The object of fast motion image was detected by improved the Faster R-CNN model.The experiments show that the Faster R-CNN method misses some small object because of the single scoring threshold when detect fast motion image.Therefore,the score of the threshold standard in the model was improved which depend on the learning situation in this paper.The duplicated or missed detection object was re-screened through the LSTM memory network for accurate object detection.The experimental results show that the recognition rate with our fusion memory network is higher than that of the existing better method Faster R-CNN.At last,these theoretical methods are applied to the object recognition of the real image,such as the high-speed rail defects detection system and the unmanned aerial vehicle(UAV)inspection system.The evaluation results demonstrate that it can recognize the image even from the high-speed motion camera.The models and the methods proposed in this dissertation are universal,and have a positive guiding for the future research of motion image restoration and object recognition.