Abstract: In various implementations, a method includes obtaining a target image having a first resolution and comprising a plurality of target pixels having a plurality of corresponding target pixel values, obtaining a guide image having a second resolution and comprising a plurality of guide pixels having a plurality of corresponding guide pixel values, and generating an enhanced target image based on the target image and the guide image, the enhanced target image having the second resolution and comprising a plurality of enhanced target pixels having a plurality of corresponding enhanced target pixel values. In various implementations, determining, for a particular upscaled target pixel, a similarity metric indicative of the similarity of a neighborhood of pixels around the particular upscaled target pixel to a neighborhood of pixels around a corresponding guide pixel, and determining, for the particular upscaled target pixel, an enhanced target pixel value based on the similarity metric.

Abstract: In various implementations, a method includes obtaining a target image having a first resolution and comprising a plurality of target pixels having a plurality of corresponding target pixel values, obtaining a guide image having a second resolution and comprising a plurality of guide pixels having a plurality of corresponding guide pixel values, and generating an enhanced target image based on the target image and the guide image, the enhanced target image having the second resolution and comprising a plurality of enhanced target pixels having a plurality of corresponding enhanced target pixel values. In various implementations, determining, for a particular upscaled target pixel, a similarity metric indicative of the similarity of a neighborhood of pixels around the particular upscaled target pixel to a neighborhood of pixels around a corresponding guide pixel, and determining, for the particular upscaled target pixel, an enhanced target pixel value based on the similarity metric.

Abstract: Techniques of reducing or eliminating artifact pixels in high dynamic range (HDR) imaging are described. One embodiment includes obtaining a first image of a scene at a first time with first exposure settings and obtaining a second image of the scene at a second time with second exposure settings that differ from the first exposure settings. The obtained images may be downsampled. The images may be compared to each other to assist with determining a number of potential artifact pixels in the scene. Depending on a relationship between the number of potential artifact pixels and a threshold value, the first image or second image may be selected as a reference image for registering the images with each other. A type of registration performed between the images may depend on which of the two images is the selected reference image. The registered images may be used to generate an HDR image.

Abstract: In various implementations, a method includes obtaining a target image having a first resolution and comprising a plurality of target pixels having a plurality of corresponding target pixel values, obtaining a guide image having a second resolution and comprising a plurality of guide pixels having a plurality of corresponding guide pixel values, and generating an enhanced target image based on the target image and the guide image, the enhanced target image having the second resolution and comprising a plurality of enhanced target pixels having a plurality of corresponding enhanced target pixel values. In various implementations, determining, for a particular upscaled target pixel, a similarity metric indicative of the similarity of a neighborhood of pixels around the particular upscaled target pixel to a neighborhood of pixels around a corresponding guide pixel, and determining, for the particular upscaled target pixel, an enhanced target pixel value based on the similarity metric.

Abstract: In various implementations a method includes obtaining a plurality of source images, stabilizing the plurality of source images to generate a plurality of stabilized images, and averaging the plurality of stabilized image to generate a synthetic long exposure image. In various implementations, stabilizing the plurality of source images includes: selecting one of the plurality of source images to serve as a reference frame; and registering others of the plurality of source images to the reference frame by applying a perspective transformation to others of the plurality of the source images.

Abstract: Image enhancement is achieved by separating image signals, e.g. YCbCr image signals, into a series of frequency bands and performing locally-adaptive noise reduction on bands below a given frequency but not on bands above that frequency. The bands are summed to develop the image enhanced signals. The YCbCr, multi-band locally-adaptive approach to denoising is able to operate independently—and in an optimized fashion—on both luma and chroma channels. Noise reduction is done based on models developed for both luma and chroma channels by measurements taken for multiple frequency bands, in multiple patches on the ColorChecker chart, and at multiple gain levels, in order to develop a simple yet robust set of models that may be tuned off-line a single time for each camera and then applied to images taken by such cameras in real-time without excessive processing requirements and with satisfactory results across illuminant types and lighting conditions.

Abstract: Techniques for de-noising a digital image using a multi-band noise filter and a unique combination of texture and chroma metrics are described. A novel texture metric may be used during multi-band filter operations on an image's luma channel to determine if a given pixel is associated with a textured/smooth region of the image. A novel chroma metric may be used during the same multi-band filter operation to determine if the same pixel is associated with a blue/not-blue region of the image. Pixels identified as being associated with a smooth blue region may be aggressively de-noised and conservatively sharpened. Pixels identified as being associated with a textured blue region may be conservatively de-noised and aggressively sharpened. By coupling texture and chroma constraints it has been shown possible to mitigate noise in an image's smooth blue regions without affecting the edges/texture in other blue objects.

Abstract: Systems, methods, and computer readable media to fuse digital images are described. In general, techniques are disclosed that use multi-band noise reduction techniques to represent input and reference images as pyramids. Once decomposed in this manner, images may be fused using novel low-level (noise dependent) similarity measures. In some implementations similarity measures may be based on intra-level comparisons between reference and input images. In other implementations, similarity measures may be based on inter-level comparisons. In still other implementations, mid-level semantic features such as black-level may be used to inform the similarity measure. In yet other implementations, high-level semantic features such as color or a specified type of region (e.g., moving, stationary, or having a face or other specified shape) may be used to inform the similarity measure.

Abstract: Systems, methods, and computer readable media to fuse digital images are described. In general, techniques are disclosed that use multi-band noise reduction techniques to represent input and reference images as pyramids. Once decomposed in this manner, images may be fused using novel low-level (noise dependent) similarity measures. In some implementations similarity measures may be based on intra-level comparisons between reference and input images. In other implementations, similarity measures may be based on inter-level comparisons. In still other implementations, mid-level semantic features such as black-level may be used to inform the similarity measure. In yet other implementations, high-level semantic features such as color or a specified type of region (e.g., moving, stationary, or having a face or other specified shape) may be used to inform the similarity measure.

Abstract: Techniques for de-noising a digital image using a multi-band noise filter and a unique combination of texture and chroma metrics are described. A novel texture metric may be used during multi-band filter operations on an image's luma channel to determine if a given pixel is associated with a textured/smooth region of the image. A novel chroma metric may be used during the the same multi-band filter operation to determine if the same pixel is associated with a blue/not-blue region of the image. Pixels identified as being associated with a smooth blue region may be aggressively de-noised and conservatively sharpened. Pixels identified as being associated with a textured blue region may be conservatively de-noised and aggressively sharpened. By coupling texture and chroma constraints it has been shown possible to mitigate noise in an image's smooth blue regions without affecting the edges/texture in other blue objects.

Abstract: Image enhancement is achieved by separating image signals, e.g. YCbCr image signals, into a series of frequency bands and performing noise reduction on bands below a given frequency but not on bands above that frequency. The bands are summed to develop the image enhanced signals. The YCbCr, multi-band approach to denoising is able to operate independently—and in an optimized fashion—on both luma and chroma channels. Noise reduction is done based on models developed for both luma and chroma channels by measurements taken for multiple frequency bands, in multiple patches on the ColorChecker chart, and at multiple gain levels, in order to develop a simple yet robust set of models that may be tuned off-line a single time for each camera and then applied to images taken by such cameras in real-time without excessive processing requirements and with satisfactory results across illuminant types and lighting conditions.

Abstract: Systems, methods, and computer readable media to capture and process high dynamic range (HDR) images when appropriate for a scene are disclosed. When appropriate, multiple images at a single—slightly underexposed—exposure value are captured (making a constant bracket HDR capture sequence) and local tone mapping (LTM) applied to each image. Local tone map and histogram information can be used to generate a noise-amplification mask which can be used during fusion operations. Images obtained and fused in the disclosed manner provide high dynamic range with improved noise and de-ghosting characteristics.

Abstract: Image enhancement is achieved by separating image signals, e.g. YCbCr image signals, into a series of frequency bands and performing noise reduction on bands below a given frequency but not on bands above that frequency. The bands are summed to develop the image enhanced signals. The YCbCr, multi-band approach to denoising is able to operate independently—and in an optimized fashion—on both luma and chroma channels. Noise reduction is done based on models developed for both luma and chroma channels by measurements taken for multiple frequency bands, in multiple patches on the ColorChecker chart, and at multiple gain levels, in order to develop a simple yet robust set of models that may be tuned off-line a single time for each camera and then applied to images taken by such cameras in real-time without excessive processing requirements and with satisfactory results across illuminant types and lighting conditions.

Abstract: Image enhancement is achieved by separating image signals, e.g. YCbCr image signals, into a series of frequency bands and performing locally-adaptive noise reduction on bands below a given frequency but not on bands above that frequency. The bands are summed to develop the image enhanced signals. The YCbCr, multi-band locally-adaptive approach to denoising is able to operate independently—and in an optimized fashion—on both luma and chroma channels. Noise reduction is done based on models developed for both luma and chroma channels by measurements taken for multiple frequency bands, in multiple patches on the ColorChecker chart, and at multiple gain levels, in order to develop a simple yet robust set of models that may be tuned off-line a single time for each camera and then applied to images taken by such cameras in real-time without excessive processing requirements and with satisfactory results across illuminant types and lighting conditions.

Abstract: Systems, methods, and computer readable media to capture and process high dynamic range (HDR) images when appropriate for a scene are disclosed. When appropriate, multiple images at a single—slightly underexposed—exposure value are captured (making a constant bracket HDR capture sequence) and local tone mapping (LTM) applied to each image. Local tone map and histogram information can be used to generate a noise-amplification mask which can be used during fusion operations. Images obtained and fused in the disclosed manner provide high dynamic range with improved noise and de-ghosting characteristics.

Abstract: A method for effectively performing local image similarity measurement is proposed. A system equipped with such a method for effectively performing an image processing task includes an image processor that performs an intermediate-results calculation procedure to calculate intermediate result values that are based upon corresponding pixels of a target patch and one or more similar patches. The image processor typically moves the target patch of the intermediate-results calculation to different locations in a raster order or some other organized order. The image processor then performs an intermediate-results combination procedure by calculating appropriate statistics of the intermediate result values to produce processed pixel values. A processor device typically controls the image processor to effectively perform the image processing tasks including, but not limited to, demosaicing and denoising.

Abstract: The present invention is directed to a method and apparatus for converting a two-dimensional image into a stereoscopic three-dimensional image. In one embodiment, a method of converting a two-dimensional image into a stereoscopic three-dimensional image comprises identifying at least one corresponding pixel from a left eye image and determining a depth and an intensity value for the each pixel within a right eye image using the at least one corresponding pixel. The depth value is stored in a right eye depth map and the intensity value is stored in the right eye image. The method further comprises inpainting at least one occluded region within the right eye image using the right eye depth map.

Abstract: A method to select an appropriate window size for local image processing uses two window sizes: a very large but impractical size and a small and practical size. Two images are generated for the desired image processing under both window sizes. The difference of these two images eliminates common non-linear nature of the local image processing and only contains low-frequency artifacts generated under a smaller window size. By taking the Fourier transform of the difference image, an approximated noise spectrum of low-frequency artifacts introduced by the small and practical window size is able to be observed. Alternatively, a frequency analysis of the smaller window size is able to be conducted by calculating the difference of cross-correlation functions and noise spectra under the same two window sizes. An appropriate window size is able to be selected based on such frequency analysis.

Abstract: Systems, methods, and computer readable media for removing noise from the luminance (luma) channel in a digital image represented in the YUV color space are described. In general, an element from the luma channel may be selected and a region about that element defined. Using a threshold that is based on the selected luma element's value, similar luma values within the defined region may be identified and combined to provide a substitute value. The substitute value may be blended with the value of the selected element within the image's luma channel. In another implementation, element values from both an image's luma and chroma channels may be used to identify similar luma values.

Abstract: A method and apparatus for multiview image generation using depth map information is described. In one embodiment, a computer-implemented method comprises converting a input image and an input depth map into a projected image and a projected depth map using values from physical pixel locations that map to projected pixel locations, wherein the projected image and the projected depth map are associated with a particular view of the input image, inpainting the projected image and the projected depth map and producing an output image in a direction of the particular view using the inpainted projected image and the inpainted projected depth map.