To take an image of object at long distance, a camera which has long-focus lens is used. However, if the focal length is long, the field of shooting becomes narrowed. Therefore, the shooting direction of the camera must be controlled. Recently, pan/tilt camera, whose direction is driven mechanically, has been used for surveillance. However, it is difficult to control quickly for long-focus camera because of weight of lens. We proposed a camera which has controllability of shooting direction without mechanical motion and high-resolution image of distant object based on computational imaging. The proposed camera is composed of ray-direction limiting optics and lens array. The direction of incident rays are limited and determined by the position of the aperture in the raydirection limiting optics, it is possible to control the shooting direction without mechanical motion. By using a lens array, a number of small optical images are formed on most area of the image sensor and digitized. Each digitized image is low-resolution because of small size. To reconstruct high-resolution image from low-resolution images, we adopted reconstruction-based super resolution technique. We did simulation under realistic conditions and confirmed the principle of the proposed camera.
Inverse quadratic problem of joint demosaicing and multiframe super-resolution(SR) was considered. Closed form solutions for different constant sub-pixel motions between frames were obtained and represented in the form of filter bank, which allows to compute solution of SR problem using adaptive filtering, where filters are selected depending on sub-pixel motion between frames. This procedure can be carried out using single iteration. For directional and non-directional parts of image corresponding directional or non-directional filters were applied. Color artifact reduction was achieved via usage of linear cross-channel regularizing term inspired by popular demosaicing methods. The framework includes motion estimation in Bayer domain, integrated noise reduction sub-algorithm, directionality estimation sub-algorithm, fallback logics and post-processing for additional color artifact reduction. Bank of filters is computed offline using specially developed compression techniques, which allows to reduce number of actually stored filters. Developed solution had shown superior results, compared to subsequent demosaicing and single channel SR and was tested on real raw images captured by cell phone camera in burst mode.
We present a maximum a posteriori (MAP) reconstruction method of increasing the pixel resolution of positron emission tomography (PET) whose typical pixel resolution is relatively low compared to that of other medical imaging modalities. We first model the underlying PET image on a finer grid and downsample it before performing forward projections to process with the observed low-resolution projection data at each iteration. We then apply a prior modeled by a linear combination of local and nonlocal regularizers to our MAP algorithm. The idea of combining the two different types of regularizers is based on our own notion that, while local regularizers are suitable for preserving fine-scale edges, non-local regularizers are suitable for preserving coarsescale edges or flat regions. Our preliminary results show that the proposed method improves the reconstruction accuracy by compromising trade-offs of the two different types of regularization on a finer grid for high-resolution reconstruction.