Abstract: The present disclosure describes neural network reduction techniques for decreasing the number of neurons or layers in a neural network. Embodiments of the method, apparatus, non-transitory computer readable medium, and system are configured to receive a trained neural network and replace certain non-linear activation units with an identity function. Next, linear blocks may then be folded to form a single block in places where the non-linear activation units were replaced by an identity function. Such techniques may reduce the number of layers in the neural network, which may optimize power and computation efficiency of the neural network architecture (e.g., without unduly influencing the accuracy of the network model).

Abstract: Methods, systems, and apparatus for combined or separate implementation of coarse-to-fine neural architecture search (NAS), two-phase block NAS, variable hardware prediction, and differential hardware design are provided and described. A variable predictor is trained, as described herein. Then, a controller or policy may be used to iteratively modify a neural network architecture along dimensions formed by neural network architecture parameters. The modification is applied to blocks (e.g., subnetworks) within the neural network architecture. In each iteration, the remainder of the neural network architecture parameters are modified and learned with a differential NAS method. The training process is performed with two-phase block NAS and incorporates a variable hardware predictor to predict power, performance, and area (PPA) parameters. The hardware parameters may be learned as well using the variable hardware predictor.

Abstract: In one example embodiment, a method includes determining at least one of integral column sums and integral row sums for pixels of an image and determining at least one of a column-wise sum of pixel values and a row-wise sum of pixel values associated with an area within the image based on at least one of the determined integral column sums and the determined integral row sums corresponding to a plurality of the pixels forming the area.

Abstract: An image processing method includes: receiving image data of a Bayer format comprising red color information, green color information, and blue color information, generating image data of a modified Bayer format by combining the green color information with the red color information and combining the green color information with the blue color information while downscaling the image data of the Bayer format to a target resolution, denoising the image data of the modified Bayer format, and generating RGB image data by demosaicing the denoised image data of the modified Bayer format.

Abstract: In one example embodiment, a method includes determining at least one of integral column sums and integral row sums for pixels of an image and determining at least one of a column-wise sum of pixel values and a row-wise sum of pixel values associated with an area within the image based on at least one of the determined integral column sums and the determined integral row sums corresponding to a plurality of the pixels forming the area.

Abstract: A noise elimination device includes a low pass filter unit, a noise filtering unit and a signal synthesis unit. The low pass filter unit generates a low frequency image signal by separating low frequency elements from an input image signal. The noise filtering unit generates a filtered image signal by suppressing noise in the low frequency image signal in horizontal and vertical directions. The signal synthesis unit generates an output image signal by combining the high frequency image signal and the filtered image signal.

Abstract: A denoising apparatus comprising an image input unit which receives pixel data including color information of pixels included in a correction target image, a denoising unit which denoises the pixel data by a weight based averaging method, wherein the weight is set to a maximum value when a difference value between a correction target block and a comparison target block in the correction target image is zero, decreases linearly to zero as the difference value increases until it reaches a threshold value, and is set to zero when the difference value is greater than or equal to the threshold value, and an image output unit which outputs the pixel data processed by the denoising unit. The denoising unit assigns a corrected weight value to at least a guaranteed number of comparison target blocks for an impulse block, where an impulse block is a correction target block for which the number of non-zero weight valued comparison target blocks is less than a predetermined guaranteed number.

Abstract: A noise elimination device includes a low pass filter unit, a noise filtering unit and a signal synthesis unit. The low pass filter unit generates a low frequency image signal by separating low frequency elements from an input image signal. The noise filtering unit generates a filtered image signal by suppressing noise in the low frequency image signal in horizontal and vertical directions. The signal synthesis unit generates an output image signal by combining the high frequency image signal and the filtered image signal.

Abstract: An image processing method includes: receiving image data of a Bayer format comprising red color information, green color information, and blue color information, generating image data of a modified Bayer format by combining the green color information with the red color information and combining the green color information with the blue color information while downscaling the image data of the Bayer format to a target resolution, denoising the image data of the modified Bayer format, and generating RGB image data by demosaicing the denoised image data of the modified Bayer format.

Abstract: A method of detecting a target in an image. The method includes receiving an image; generating a plurality of scaled-down images based on the received image; generating integral column images of each of the plurality of scaled images by calculating integral values of pixels column by column; selecting and classifying a plurality of windows of the integral column images according to a feature arithmetic operation based on a recursive column calculation; and detecting the target on the basis of the classification results for the plurality of windows.

Abstract: A denoising apparatus comprising an image input unit which receives pixel data including color information of pixels included in a correction target image, a denoising unit which denoises the pixel data by a weight based averaging method, wherein the weight is set to a maximum value when a difference value between a correction target block and a comparison target block in the correction target image is zero, decreases linearly to zero as the difference value increases until it reaches a threshold value, and is set to zero when the difference value is greater than or equal to the threshold value, and an image output unit which outputs the pixel data processed by the denoising unit. The denoising unit assigns a corrected weight value to at least a guaranteed number of comparison target blocks for an impulse block, where an impulse block is a correction target block for which the number of non-zero weight valued comparison target blocks is less than a predetermined guaranteed number.

Abstract: At least one example embodiment provides for a noise reducing system for an image sensor having a pixel array, the noise reducing system includes a pattern matcher configured to receive a first pixel value of a first pixel in the pixel array and output a reduced noise pixel value of the first pixel based on a comparison of a pixel value pattern including the first pixel value and at least another pixel value pattern in the pixel array.

Abstract: A sensing circuit based on an application-specific integrated circuit (ASIC) sensor which includes a sensor portion and a processor portion which are integrated on an ASIC. The sensor portion outputs raw output in response to a stimulus. The output of the sensor portion is processed by the processor portion. The sensor portion and the processor portion together form at least two blocks which are configurable together by interconnections in two or more ways to produce differentiated sensing products.

Abstract: A method of operating an image sensor apparatus is provided. In this method, first digital image data is generated based on a first exposure data signal, wherein the first exposure data signal is indicative of an exposure level of the image sensor during a first exposure period. The first digital image data is stored in a storage circuit. Second digital image data is generated based on a second exposure data signal, wherein the second exposure data signal is indicative of an exposure level of the image sensor during a second exposure period. The second exposure period is different from the first exposure period. The second digital image data associated with the image sensor is selectively stored based on a value of the first digital image data associated with the image sensor.