Abstract: An image processing method is provided. The method includes acquiring a video. The method includes using an object detection engine to detect a person in the video. The object detection engine is integrated with an image signal processing pipeline. The method includes transmitting the video over a network. The method includes determining that the detected person has moved less than a pre-set distance. The method includes, responsive to the determining, pausing transmission of the video. An embedded image processor including an object detection engine is also provided.

Abstract: Devices, methods, and non-transitory program storage devices are disclosed herein to provide improved multi-spectral image processing techniques for generating an enhanced output image, the techniques comprising: obtaining an N-channel (e.g., multispectral) input image; determining fusion weights and fallback weights (e.g., relative intensity weights) for each of the N-channels of the input image; blending the fusion and fallback weights based on an amount of gradient information to generate blended weights; modulating the blended weights for a plurality of frequency band representations of the input image; applying the modulated blended weights to the corresponding frequency band representations of the input image to generate a plurality of output image frequency band representations; producing an output luma image, based on the plurality of output image frequency band representations; and generating an output RGB image, based on the output luma image, which may then, e.g.

Abstract: One limitation of traditional imaging systems is that they are only programmed to correct for a single color of illuminant in a scene. In multi-illuminant scenes, the detected illuminant color may correspond to some mixture of scene illuminants. This may lead to incomplete color correction, wherein, e.g., the dominant illuminant is corrected for but the color cast caused by secondary illuminants is still visible, or an at least partially visible color cast remains from multiple of the scene illuminants. Thus, the techniques disclosed herein comprise: obtaining an image of a scene; generating an illumination map for the obtained image; dividing the values in the illumination map to determine a number of estimated illuminant regions, wherein each region corresponds to at least one estimated illuminant present in the captured scene; estimating a white point for each region; and applying white balancing operations, based on the estimated white points for each region.

Abstract: Devices, methods, and non-transitory program storage devices for spatiotemporal image noise reduction are disclosed, comprising: maintaining an accumulated image in memory; and obtaining a first plurality of multispectral images (e.g., RGB-IR images). For each image in the first plurality of multispectral images, the method may: calculate a multispectral guide image for the current image; calculate blending weights for the current image; apply the calculated blending weights to each channel of the current image to generate a denoised current image; and update the accumulated image based on pixel differences between the denoised current image and the accumulated image. In some embodiments, additional images (e.g., the accumulated image and/or other images captured prior to or after a given current image) may also be included in the denoising operations for a given current image. Finally, the method may generate a denoised output image for each input image, based on the updated accumulated image.

Abstract: A machine learning system for processing image data obtained from an image sensor is provided. The system includes a front end comprising one or more hard-coded filters, each of the one or more hard-coded filters being arranged to perform a set task. The system includes a neural network arranged to receive and process output from the front end. The one or more hard-coded filters include one or more hard-coded noise compensation filters that are hard-coded to compensate for a noise profile of the image sensor from which the image data is obtained. A method of processing image data in a machine learning system is also provided. A system for processing image data is provided.

Abstract: Devices, methods, and non-transitory program storage devices are disclosed herein to provide improved multi-spectral image processing techniques for generating an enhanced output image, the techniques comprising: obtaining an N-channel (e.g., multispectral) input image; determining fusion weights and fallback weights (e.g., relative intensity weights) for each of the N-channels of the input image; blending the fusion and fallback weights based on an amount of gradient information to generate blended weights; modulating the blended weights for a plurality of frequency band representations of the input image; applying the modulated blended weights to the corresponding frequency band representations of the input image to generate a plurality of output image frequency band representations; producing an output luma image, based on the plurality of output image frequency band representations; and generating an output RGB image, based on the output luma image, which may then, e.g.

Abstract: One limitation of traditional imaging systems is that they are only programmed to correct for a single color of illuminant in a scene. In multi-illuminant scenes, the detected illuminant color may correspond to some mixture of scene illuminants. This may lead to incomplete color correction, wherein, e.g., the dominant illuminant is corrected for but the color cast caused by secondary illuminants is still visible, or an at least partially visible color cast remains from multiple of the scene illuminants. Thus, the techniques disclosed herein comprise: obtaining an image of a scene; generating an illumination map for the obtained image; dividing the values in the illumination map to determine a number of estimated illuminant regions, wherein each region corresponds to at least one estimated illuminant present in the captured scene; estimating a white point for each region; and applying white balancing operations, based on the estimated white points for each region.

Abstract: Devices, methods, and non-transitory program storage devices for spatiotemporal image noise reduction are disclosed, comprising: maintaining an accumulated image in memory; and obtaining a first plurality of multispectral images (e.g., RGB-IR images). For each image in the first plurality of multispectral images, the method may: calculate a multispectral guide image for the current image; calculate blending weights for the current image; apply the calculated blending weights to each channel of the current image to generate a denoised current image; and update the accumulated image based on pixel differences between the denoised current image and the accumulated image. In some embodiments, additional images (e.g., the accumulated image and/or other images captured prior to or after a given current image) may also be included in the denoising operations for a given current image. Finally, the method may generate a denoised output image for each input image, based on the updated accumulated image.

Abstract: There is provided a method for automatically altering a digital video stream including multiple video input frames, to automatically obtain output frames with a target composition, in which frame metadata relating to objects in one or more of the video input frames is analyzed on a frame-by-frame basis and used by a processor to automatically alter one or more output frames to be more similar to, or to match, the target composition, wherein cropping is performed in 3D. A related system and a related computer program product are also provided.

Abstract: A method including receiving an indication that a first classifier has identified that an image includes an object of a predetermined class of objects. Image data that relates to the image is processed using a second classifier with a first training state, which determines whether the image data includes the object of the predetermined class of objects. In response to the determining, data relating to the image data is transmitted to a remote system. Update data relating to the transmitted data is received from the remote system. The training state of the second classifier is updated to a second training state in response to the update data such that the second classifier with the second training state would make a different determination of whether future image data similar to the image data includes an object of the predetermined class of objects than the second classifier with the first training state.

Abstract: According to an aspect of the present disclosure, there is provided a method of image processing. The method comprises receiving image data comprising a set of feature vectors of a first dimensionality, the feature vectors corresponding to a class of objects. A variable projection is applied to each feature vector in the set of feature vectors to generate a set of projected vectors of a second dimensionality. The method then comprises processing the set of projected vectors to generate a model for the class of objects. A projection is applied to the model to generate an object classification model, of the first dimensionality, for the class of objects.

Abstract: A computer vision process in which (i) a track record for one or more detected objects is created, the track record defining metadata or parameters for that detected object; and (ii) that track record is then linked to one or more higher-level track records for one or more higher-level objects.

Abstract: A machine learning system for processing image data obtained from an image sensor is provided. The system includes a front end comprising one or more hard-coded filters, each of the one or more hard-coded filters being arranged to perform a set task. The system includes a neural network arranged to receive and process output from the front end. The one or more hard-coded filters include one or more hard-coded noise compensation filters that are hard-coded to compensate for a noise profile of the image sensor from which the image data is obtained. A method of processing image data in a machine learning system is also provided. A system for processing image data is provided.

Abstract: A method including receiving an indication that a first classifier has identified that an image includes an object of a predetermined class of objects. Image data that relates to the image is processed using a second classifier with a first training state, which determines whether the image data includes the object of the predetermined class of objects. In response to the determining, data relating to the image data is transmitted to a remote system. Update data relating to the transmitted data is received from the remote system. The training state of the second classifier is updated to a second training state in response to the update data such that the second classifier with the second training state would make a different determination of whether future image data similar to the image data includes an object of the predetermined class of objects than the second classifier with the first training state.

Abstract: There is provided a method for estimating a distance of an object detected by an image sensor. Multiple detections are performed automatically to detect features of an object and to estimate the object proportions, which are then used to relate to additional measurements such as the distance of the object from the image sensor. The method detects human and non-human objects. The method uses available anthropometry tables. The method takes into account the image sensor optical aberrations such as lens distortion. A related system and a related computer program product are also provided.

Abstract: An image processing method is provided. The method includes acquiring a video. The method includes using an object detection engine to detect a person in the video. The object detection engine is integrated with an image signal processing pipeline. The method includes transmitting the video over a network. The method includes determining that the detected person has moved less than a pre-set distance. The method includes, responsive to the determining, pausing transmission of the video. An embedded image processor including an object detection engine is also provided.

Abstract: A method to frame a video stream is provided. A video stream is captured, then motion of the camera, object presence in the video stream, and position of the object are detected. A framed video stream is generated using a framing in dependence on the motion of the camera and the position of the object. A system for framing a video stream is also provided, wherein the system comprises a camera and a processing unit, and the processing unit comprises at least one processor and at least on memory including computer program instructions to execute a method to frame a video stream.

Abstract: The invention relates to feature extraction technique based on edge extraction. It can be used in computer vision systems, including image/facial/object recognition systems, scene interpretation, classification and captioning systems. A model or profile of the noise in the sensor is used to improve feature extraction or object detection on an image from a sensor.

Abstract: A method to analyze an image and determine whether to output image data associated with an area of the image is provided. An object detection algorithm using training image data to detect an object based at least in part on a similarity of appearance of image data to data derived from the training image data is provided. Weakly detected objects are classified based on characteristics associated with the weakly detected object and may be added to the training image dataset for use in further training of the object detection algorithm. The object detection algorithm is trained with a revised dataset, the revised dataset being updated with data generated by the object detection algorithm.

Abstract: There is provided a method for automatically altering a digital video stream including multiple video input frames, to automatically obtain output frames with a target composition, in which frame metadata relating to objects in one or more of the video input frames is analyzed on a frame-by-frame basis and used by a processor to automatically alter one or more output frames to be more similar to, or to match, the target composition, wherein cropping is performed in 3D. A related system and a related computer program product are also provided.