Abstract: The present disclosure provides a video-based behavior recognition device, comprising a synthesized channel frame provision unit generating highlight information by comparing channel frames corresponding to the respective channels among a plurality of channels and synthesizing the channel frames and the highlight information to provide a synthesized channel frame, a neural network unit providing a middle frame on the basis of the synthesized channel frame and a multi-frame convolution neural network, and a behavior recognition result provision unit providing a behavior recognition result on the basis of the middle frame and a weighted value generated according to the middle frame, and a method operating thereof.

Abstract: A color fringe is corrected by detecting a transition region that includes pixels adjacent in a linear direction. A color difference distribution in the transition region is modeled by a logistic function. Pixel color values in the transition region are corrected using the logistic function to maximize a correlation between a correction color and a reference color with respect to the transition region. Color distortion such as color fringes is corrected without corrupting the original colors of the image by modeling the color difference by the logistic function while maximizing the correlation using information of the undistorted region. A calculation cost is reduced by reducing the number of the parameters required to optimize the logistic function.

Abstract: An image haze removing apparatus includes at least a transmission coefficient generator, a low pass filter, a particle size parameter estimator, a transmission coefficient divider and a real image estimator. The transmission coefficient generator generates a transmission coefficient based on a captured image. The low pass filter outputs a low frequency image, which includes a low frequency component of the captured image. The particle size parameter estimator estimates a particle size parameter based on the low frequency image. The transmission coefficient divider divides the transmission coefficient based on the particle size parameter to generate first to third divided transmission coefficients, which correspond to red, green and blue data of the captured image, respectively. The real image estimator estimates a real image by applying the first to third divided transmission coefficients to the red, green and blue data of the captured image, respectively.

Abstract: A device for removing haze from an image includes an image decomposer that decomposes a near-infrared (NIR) image to generate an NIR detail layer image and decomposes an RGB image to generate an RGB detail layer image and an RGB base layer image, a weight generator that generates a mixing weight value based on a similarity between high frequency (HF) components of the NIR image and the RGB image, a detail layer mixer that mixes the NIR detail layer image and the RGB detail layer image based on the mixing weight value to generate a mixed RGB detail layer image, a base layer dehazer that removes haze from the RGB base layer image to generate a compensated RGB base layer image, and an adder that adds the mixed RGB detail layer image and the compensated RGB base layer image to generate an output RGB image.

Abstract: A color fringe is corrected by detecting a transition region that includes pixels adjacent in a linear direction. A color difference distribution in the transition region is modeled by a logistic function. Pixel color values in the transition region are corrected using the logistic function to maximize a correlation between a correction color and a reference color with respect to the transition region. Color distortion such as color fringes is corrected without corrupting the original colors of the image by modeling the color difference by the logistic function while maximizing the correlation using information of the undistorted region. A calculation cost is reduced by reducing the number of the parameters required to optimize the logistic function.

Abstract: An image processing device includes a guided image filtering circuit and an image summation circuit. The guided image filtering circuit is configured to receive an input image including a green channel, a red channel and a blue channel, and to generate a corrected red channel and a corrected blue channel by applying a guided filter with respect to the red channel and the blue channel based on the green channel. The image summation circuit is configured to generate an output image by summing the green channel, the corrected red channel, and the corrected blue channel.

Abstract: A device for removing haze from an image includes an image decomposer that decomposes a near-infrared (NIR) image to generate an NIR detail layer image and decomposes an RGB image to generate an RGB detail layer image and an RGB base layer image, a weight generator that generates a mixing weight value based on a similarity between high frequency (HF) components of the NIR image and the RGB image, a detail layer mixer that mixes the NIR detail layer image and the RGB detail layer image based on the mixing weight value to generate a mixed RGB detail layer image, a base layer dehazer that removes haze from the RGB base layer image to generate a compensated RGB base layer image, and an adder that adds the mixed RGB detail layer image and the compensated RGB base layer image to generate an output RGB image.

Abstract: An image haze removing apparatus includes at least a transmission coefficient generator, a low pass filter, a particle size parameter estimator, a transmission coefficient divider and a real image estimator. The transmission coefficient generator generates a transmission coefficient based on a captured image. The low pass filter outputs a low frequency image, which includes a low frequency component of the captured image. The particle size parameter estimator estimates a particle size parameter based on the low frequency image. The transmission coefficient divider divides the transmission coefficient based on the particle size parameter to generate first to third divided transmission coefficients, which correspond to red, green and blue data of the captured image, respectively. The real image estimator estimates a real image by applying the first to third divided transmission coefficients to the red, green and blue data of the captured image, respectively.

Abstract: In an image processing device, a color transformation circuit is configured to generate a color difference based on a first image of normal exposure and a second image of under exposure. A color fringe detection circuit is configured to detect color fringe in the first image, and to generate color fringe detection information based on the detected color fringe. A weight map formation circuit is configured to generate a weight map based on the color fringe detection information and the color difference. An image fusion circuit is configured to generate a third image, in which the detected color fringe is suppressed, based on the color difference and the weight map.

Abstract: An image processing device includes a guided image filtering circuit and an image summation circuit. The guided image filtering circuit is configured to receive an input image including a green channel, a red channel and a blue channel, and to generate a corrected red channel and a corrected blue channel by applying a guided filter with respect to the red channel and the blue channel based on the green channel. The image summation circuit is configured to generate an output image by summing the green channel, the corrected red channel, and the corrected blue channel.

Abstract: In an image processing device, a color transformation circuit is configured to generate a color difference based on a first image of normal exposure and a second image of under exposure. A color fringe detection circuit is configured to detect color fringe in the first image, and to generate color fringe detection information based on the detected color fringe. A weight map formation circuit is configured to generate a weight map based on the color fringe detection information and the color difference. An image fusion circuit is configured to generate a third image, in which the detected color fringe is suppressed, based on the color difference and the weight map.

Abstract: Methods of correcting saturated pixel data in an image sensor are provided. A method of correcting saturated pixel data in an image sensor includes determining a weight function. The weight function indicates a correlation between color values of saturated pixels and color values of neighboring pixels. The saturated pixels are among a plurality of pixels which have a color value greater than a saturation threshold value. The neighboring pixels are among the plurality of pixels that are proximate to each of the saturated pixels. The method includes determining weight values of a neighboring pixels that are proximate to a first saturated pixel using the weight function. The method includes determining a weighted average value of the color values of each of the neighboring pixels using the weight values. The method includes correcting the color value of the first saturated pixel to the weighted average value.

Abstract: Methods of correcting saturated pixel data in an image sensor are provided. A method of correcting saturated pixel data in an image sensor includes determining a weight function. The weight function indicates a correlation between color values of saturated pixels and color values of neighboring pixels. The saturated pixels are among a plurality of pixels which have a color value greater than a saturation threshold value. The neighboring pixels are among the plurality of pixels that are proximate to each of the saturated pixels. The method includes determining weight values of a neighboring pixels that are proximate to a first saturated pixel using the weight function. The method includes determining a weighted average value of the color values of each of the neighboring pixels using the weight values. The method includes correcting the color value of the first saturated pixel to the weighted average value.

Abstract: An image processing method and an image processing apparatus for removing noise from an image are disclosed. A provided image processing method includes: dividing an input image into a luminance signal and a chrominance signal; removing noise from the luminance signal; restoring luminance signal present in the noise removed from the luminance signal; removing noise from the chrominance signal; and combining the luminance signal and the chrominance signal from which the noises are removed. Accordingly, an image of which an edge component is well preserved and a degree of color noise is low is generated not only in a general environment but also in a low light level and high sensitivity environment having a large amount of noise.

Abstract: In image processing, a color fringe is removed. The processing includes selecting a maximum gradient magnitude among gradient magnitudes for each of color components in an image, calculating a boundary of a dilated near-saturation region in the image, detecting a transition region in the image according to the maximum gradient magnitude and the dilated near-saturation region, and removing a color fringe from the transition region.

Abstract: Provided is a method for assessing image quality using quantization codes, which includes: filtering an original image and a distorted image; generating phase quantization codes from the filtering result; calculating a Hamming difference between the phase quantization code of the original image and the phase quantization code of the distorted image; and assessing image quality of the distorted image by using the calculated Hamming difference. According to the present disclosure, since pixel values of the original image and the distorted image are mapped onto a quantized complex plane and then binary code operation is performed, it is possible to easily implement image quality assessing hardware and also ensure excellent image quality assessing performance.

Abstract: An image processing method and image processing apparatus for removing noise from an image. The image processing method includes: dividing an input image into a high frequency component and a low frequency component; removing noise from the high frequency component; and combining the low frequency component and the high frequency component from which noise is removed to form an output image. Accordingly, an image in which edge components are well preserved and a degree of color noise is reduced may be generated not only in a general environment but also in a low light level and highly sensitive environment having strong noise.

Abstract: An image processing apparatus and an image processing method, the method including: providing one or more low dynamic range images having different exposure levels; classifying each of the one of more low dynamic range images as a saturation region, an under-saturation region, and a non-saturation region with respect to an intensity value; generating feature maps indicating a similarity between patterns of the non-saturation regions; generating final feature maps converting pixels included in the feature maps into moving pixels and non-moving pixels; defining moving regions and non-moving regions included in the one or more low dynamic range images; removing noise from the moving regions and the non-moving regions by using a first noise reduction filter for the moving regions and a second noise reduction filter for the non-moving regions; and generating a radiance map corresponding to the one or more low dynamic range images.

Abstract: In image processing, a color fringe is removed. The processing includes selecting a maximum gradient magnitude among gradient magnitudes for each of color components in an image, calculating a boundary of a dilated near-saturation region in the image, detecting a transition region in the image according to the maximum gradient magnitude and the dilated near-saturation region, and removing a color fringe from the transition region.

Abstract: An image interpolating method and apparatus, in which horizontal and vertical Differences of Absolute Inter-channel Differences (DAIDs) are calculated from a color filter array (CFA) image, and an unknown pixel is interpolated in horizontal and vertical directions estimated in consideration of the DAIDs of R, G, and B pixels. Therefore, the image interpolating method and apparatus provide a large Peak Signal and Noise Ratio (PSNR), a Structural Similarity (SSIM), and high visual quality images.