| Computational optical imaging technology is a comprehensive application of modern optical system design methods,image sensing technology and image processing technology.It can achieve a qualitative breakthrough in the multi-dimensional,multi-scale and resolution of visual information,and then observe the invisible,incomplete and invisible scene information of traditional optical imaging system.In recent years,how to use image processing technology to improve the image quality of computational optical system has become a hot issue in the research of complex optical scene with low contrast,low resolution and low illumination.For this reason,this paper has carried out three aspects of research work: polarization imaging defogging enhancement processing,infrared night vision image super-resolution reconstruction and underwater dim and small target detection and recognition.The specific research contents are as follows:(1)For dense haze scenes,in order to enhance the image contrast and enhance the detection distance of the imaging system,a four cameras real-time polarization differential imaging system with common optical axis is proposed based on Stokes vector model.Four polarization component images can be obtained in real time by integrating the polarizers with the angles of 0 °,45 °,90 ° and 135 ° with the incident light.In the field of polarization component image registration,a real-time registration algorithm for polarization component image is proposed.In the field of polarization component image registration,a real-time registration algorithm is proposed.Firstly,the overlapped area of polarization component image is calibrated at pixel level by using high-precision four-dimensional calibration platform,and the image features of overlapped area are extracted by speeded up robust features(SURF)detector.Then,we use fast approximate nearest neighbors(FANN)search algorithm to get the initial matching point pairs.By sorting and selecting the Euclidean geometric distance of the matching points to the feature vector,we can keep the better matching point pairs.Finally,the sampling process of interior point set is optimized globally based on progressive sample consensus(PROSAC)algorithm to achieve more accurate spatial transformation parameter estimation.In the aspect of polarization synthesis image video fusion,a video fusion method based on the frame of mapping merge is proposed,and the video is visually enhanced by using the contrast limited adaptive histgram equalization(CLAHE)algorithm.The experimental results show that this method can obtain high contrast and high definition images in real time without losing the spatial resolution of the imaging system.(2)For low illumination scenes at night,in order to improve the resolution of infrared imaging system and remove noise interference,a super-resolution reconstruction algorithm based on microscanning imaging is proposed.Firstly,the reference frame is constructed by bilinear interpolation.Then,the four parameters affine transformation model is used to estimate the motion parameters of the sequence image.Then,a point spread function(PSF)operator based on edge detection is proposed for convex set projection and gray level correction.Finally,a k-means singular value decomposition(K-SVD)dictionary learning algorithm based on sparse threshold is proposed for image denoising.The experimental results show that this method can recover the edge details of the image and suppress the noise to a certain extent.In addition,an image super-resolution reconstruction algorithm based on compressed coding imaging is proposed and applied to coding aperture infrared night vision imaging system.Firstly,an aperture encoder is added to the spectrum plane to filter the redundant information of the image,and the sampled image signal is coded and compressed by Fourier transform.Then,the sparse representation of signals in Fourier transform domain is realized by using the observation matrix generated by coded aperture.Finally,a global optimization gradient projection parallel acceleration algorithm combining two-point gradient method and adaptive non-monotone line search strategy is proposed for super-resolution reconstruction of sparse signals.The experimental results show that the method can reconstruct high resolution image information with far less data than the original signal,and improve the efficiency of image reconstruction.In addition,a super-resolution reconstruction algorithm based on compression coding image is proposed.Firstly,the image is sparse represented by Fourier transform.Then,a global optimization gradient projection parallel acceleration algorithm based on the adaptive non monotone line search strategy is proposed to complete the super-resolution reconstruction of sparse signals.The experimental results show that this method can efficiently reconstruct a high-resolution image with less data and suppress image noise.(3)For turbid underwater scenes,in order to improve the detection and recognition accuracy of dim and small targets in turbid media,an image edge detection method based on ant colony optimization(ACO)and reinforcement learning is proposed.Firstly,the reinforcement learning idea is integrated into the artificial ant movement,and a variable sensing radius strategy is proposed to calculate the transfer probability of each pixel,which is different from the traditional method which uses a fixed number of domain pixels to calculate the gradient.Then,a dual population strategy is introduced to control the movement direction of ants,which makes the search process take into account both global and local search capabilities.The experimental results show that this method can detect the edge of small and weak underwater target with high precision.In addition,an underwater image target detection and recognition method based on deep codec network is proposed.Firstly,the image features extracted by alexnet convolution operation are supplemented by deconvolution operation,and an encoding and decoding network structure based on AlexNet(ED-AlexNet)is proposed.Then,the problem of data starvation is solved by migration learning,and the training sample set is expanded based on affine transformation and other operations.The experimental results show that the method has high recognition accuracy for small and weak underwater targets. |
PDF Full Text Request