3D Reconstruction for Motion Blurred Images Using Deep Learning-Based Intelligent Systems

The 3D reconstruction using deep learning-based intelligent systems can provide great help for measuring an individual’s height and shape quickly and accurately through 2D motion-blurred images.Generally,during the acquisition of images in real-time,motion blur,caused by camera shaking or human motion,appears.Deep learning-based intelligent control applied in vision can help us solve the problem.To this end,we propose a 3D reconstruction method for motion-blurred images using deep learning.First,we develop a BF-WGAN algorithm that combines the bilateral filtering(BF)denoising theory with a Wasserstein generative adversarial network(WGAN)to remove motion blur.The bilateral filter denoising algorithm is used to remove the noise and to retain the details of the blurred image.Then,the blurred image and the corresponding sharp image are input into the WGAN.This algorithm distinguishes the motion-blurred image from the corresponding sharp image according to the WGAN loss and perceptual loss functions.Next,we use the deblurred images generated by the BFWGAN algorithm for 3D reconstruction.We propose a threshold optimization random sample consensus(TO-RANSAC)algorithm that can remove the wrong relationship between two views in the 3D reconstructed model relatively accurately.Compared with the traditional RANSAC algorithm,the TO-RANSAC algorithm can adjust the threshold adaptively,which improves the accuracy of the 3D reconstruction results.The experimental results show that our BF-WGAN algorithm has a better deblurring effect and higher efficiency than do other representative algorithms.In addition,the TO-RANSAC algorithm yields a calculation accuracy considerably higher than that of the traditional RANSAC algorithm.

Blind-restoration-based blind separation method for permuted motion blurred images

A novel single-channel blind separation algorithm for permuted motion blurred images is proposed by using blind restoration in this paper. Both the motion direction and the length of the point spread function （PSF） are estimated by Radon transformation and extrema a detection. Using the estimated blur parameters, the permuted image is restored by performing the L-R blind restoration method. The permutation mixing matrices can be accurately estimated by classifying the ringing effect in the restored image, thereby the source images can be separated. Simulation results show a better separation efficiency for the permuted motion blurred image with various permutation operations. The proposed algorithm indicates a better performance on the robustness against Gaussian noise and lossy JPEG compression.

Numericals for total variation-based reconstruction of motion blurred images

In this paper image with horizontal motion blur, vertical motion blur and angled motion blur are considered. We construct several difference schemes to the highly nonlinear term △↓.（△↓u/√｜△↓｜^2＋β） of the total variation-based image motion deblurring problem. The large nonlinear system is linearized by fixed point iteration method. An algebraic multigrid method with Krylov subspace acceleration is used to solve the corresponding linear equations as in [7]. The algorithms can restore the image very well. We give some numerical experiments to demonstrate that our difference schemes are efficient and robust.

Radon transform-based motion blurred silkworm pupa image restoration

As for machine vision-based intelligent system in the application of discriminating and sorting the sex of silkworm pupae,the tail gonad was the unique physiological feature.However,motion blur,resulting from the live silkworm pupa’s writhing motion at the moment of capturing image,could lose textures and structures(such as edge and tail gonad etc.)dramatically,which casted great challenges for sex identification.To increase the image quality and relieve the difficulty of discrimination caused by motion blur,an effective approach that including three stages was proposed in this work.In the image prediction stage,first sharp edges were acquired by using filtering techniques.Then the initial blur kernel was computed with Gaussian prior.The coarse version latent image was deconvoluted in the Fourier domain.In the kernel refinement stage,the Radon transform was applied to estimate the accurate kernel.In the final restoration step,a TV-L1 deconvolution model was carried out to obtain a better result.The experimental results showed that benefiting from the prediction step and kernel refinement step,the kernel was more accurate and the recovered image contained much more textures.It revealed that the proposed method was useful in removing the motion blur.Furthermore,the method could also be applied to other fields.

Camera Independent Motion Deblurring in Videos Using Machine Learning

In this paper, we will be looking at our efforts to find a novel solution for motion deblurring in videos. In addition, our solution has the requirement of being camera-independent. This means that the solution is fully implemented in software and is not aware of any of the characteristics of the camera. We found a solution by implementing a Convolutional Neural Network-Long Short Term Memory (CNN-LSTM) hybrid model. Our CNN-LSTM is able to deblur video without any knowledge of the camera hardware. This allows it to be implemented on any system that allows the camera to be swapped out with any camera model with any physical characteristics.

Method for predicting motion artifacts of plasma display panels

A simulation method is proposed to predict the motion artifacts of plasma display panels （PDPs）. The method simulates the behavior of the human vision system when perceiving moving objects. The simulation is based on the measured temporal light properties of the display for each gray level and each phosphor. Both the effect of subfield arrangement and phosphor decay are involved. A novel algorithm is proposed to improve the calculation speed. The simulation model manages to predict the appearance of the motion image perceived by a human with a still image. The results are validated by a set of perceptual evaluation experiments. This rapid and accurate prediction of motion artifacts enables objective characterization of the PDP performance in this aspect.

MEASUREMENT OF ANGULAR VIBRATION AMPLITUDE BY ACTIVELY BLURRED IMAGES

A novel motion-blur-based method for measuring the angular amplitude of a high-frequency rotational vibration is schemed. The proposed approach combines the active vision concept and the mechanism of motion-from-blur, generates motion blur on the image plane actively by extending exposure time, and utilizes the motion blur information in polar images to estimate the angular amplitude of a high-frequency rotational vibration. This method obtains the analytical results of the angular vibration amplitude from the geometric moments of a motion blurred polar image and an unblurred image for reference. Experimental results are provided to validate the presented scheme.

Research on single image super-resolution based on very deep super-resolution convolutional neural network

Single image super-resolution(SISR)is a fundamentally challenging problem because a low-resolution(LR)image can correspond to a set of high-resolution(HR)images,while most are not expected.Recently,SISR can be achieved by a deep learning-based method.By constructing a very deep super-resolution convolutional neural network(VDSRCNN),the LR images can be improved to HR images.This study mainly achieves two objectives:image super-resolution(ISR)and deblurring the image from VDSRCNN.Firstly,by analyzing ISR,we modify different training parameters to test the performance of VDSRCNN.Secondly,we add the motion blurred images to the training set to optimize the performance of VDSRCNN.Finally,we use image quality indexes to evaluate the difference between the images from classical methods and VDSRCNN.The results indicate that the VDSRCNN performs better in generating HR images from LR images using the optimized VDSRCNN in a proper method.

Blind image deconvolution for single motion-blurred image

Motion blur due to camera shake during exposure is one of the most common reasons of image degradation, which usually reduces the quality of photographs seriously. Based on the statistical properties of the natural image＇s gradient and the blur kernel, a blind deconvolution algorithm is proposed to restore the motion-blurred image caused by camera shake, adopting the variational Bayesian estimation theory. In addition, the ring effect is one problem that is not avoided in the process of image deconvolution, and usually makes the visual effect of the restored image badly. So a dering method is put forward based on the sub-region detection and fuzzy filter. Tested on the real blurred photographs, the experimental results show that the proposed algorithm of blind image deconvolution can remove the camera-shake motion blur from the degraded image effectively, and can eliminate the ring effect better, while preserve the edges and details of the image well.

Multi-Dimension Support Vector Machine Based Crowd Detection and Localisation Framework for Varying Video Sequences

In this paper, we propose a novel method for anomalous crowd behaviour detection and localization with divergent centers in intelligent video sequence through multiple SVM (support vector machines) based appearance model. In multi-dimension SVM crowd detection, many features are available to track the object robustly with three main features which include 1) identification of an object by gray scale value, 2) histogram of oriented gradients (HOG) and 3) local binary pattern (LBP). We propose two more powerful features namely gray level co-occurrence matrix (GLCM) and Gaber feature for more accurate and authenticate tracking result. To combine and process the corresponding SVMs obtained from each features, a new collaborative strategy is developed on the basis of the confidence distribution of the video samples which are weighted by entropy method. We have adopted subspace evolution strategy for reconstructing the image of the object by constructing an update model. Also, we determine reconstruction error from the samples and again automatically build an update model for the target which is tracked in the video sequences. Considering the movement of the targeted object, occlusion problem is considered and overcome by constructing a collaborative model from that of appearance model and update model. Also if update model is of discriminative model type, binary classification problem is taken into account and overcome by collaborative model. We run the multi-view SVM tracking method in real time with subspace evolution strategy to track and detect the moving objects in the crowded scene accurately. As shown in the result part, our method also overcomes the occlusion problem that occurs frequently while objects under rotation and illumination change due to different environmental conditions.

Research on Motion Estimation Algorithm of Star Point Based on Nonlinear Gaussian Fitting

The energy distribution model of motion blurred star point is analyzed.The distribution of the star point approximates to a two-dimensional(2 D) Gaussian distribution under degeneration.Two multi-parameter nonlinear Gaussian fitting methods(GFMs) are proposed,and the relationship between fitting parameters and motion blur parameters is analyzed.Estimation of the parameters of motion blur by fitting parameters is calculated to realize the error compensation of the motion blur.The simulation results show the effectiveness and accuracy.

Real-time motion deblurring algorithm with robust noise suppression

In an image restoration process,to obtain good results is challenging because of the unavoidable existence of noise even if the blurring information is already known.To suppress the deterioration caused by noise during the image deblurring process,we propose a new deblurring method with a known kernel.First,the noise in the measurement process is assumed to meet the Gaussian distribution to fit the natural noise distribution.Second,the first and second orders of derivatives are supposed to satisfy the independent Gaussian distribution to control the non-uniform noise.Experimental results show that our method is obviously superior to the Wiener filter,regularized filter,and Richardson-Lucy(RL) algorithm.Moreover,owing to processing in the frequency domain,it runs faster than the other algorithms,in particular about six times faster than the RL algorithm.