Abstract: A system for monitored tomographic reconstruction, comprising: an x-ray generator configure to generate x-ray beams for scanning an object; detectors configured to capture a plurality of projections for each scan; at least one hardware processor; and one or more software modules that, when executed by the at least one hardware processor, receive the plurality of projections from the detectors and as each of the plurality of projections is received, generate a partial reconstruction, and make a stopping decision with respect to whether or not another projection should be obtained based on a stopping problem and that defines when a reconstructed image quality is sufficient with respect to the expended cost as determined by a stopping rule.

Abstract: The computation of local feature descriptors for image matching in computer vision can be computationally expensive for real-time on-device applications. Accordingly, disclosed embodiments speed up such computations by precomputing values for gradient maps, and storing them in lookup tables indexed by partial derivatives. In addition, certain embodiments introduce global smoothing, and optionally, global gradient maps. In an embodiment that eliminates all floating-point operations, arctangents can be precomputed for a fixed number of angles, and quantization can be performed when computing the local feature descriptors.

Abstract: Given their limited computational resources, mobile or embedded devices may not be capable of operating a full-precision convolutional neural network (CNN). Thus, quantized neural networks (QNNs) may be used in place of a full-precision CNN. For example, 8-bit QNNs have similar accuracy to full-precision CNNs. While lower-bit QNNs, such as 4-bit QNNs, are faster and more computationally efficient than 8-bit QNNs, they are also significantly less accurate. Accordingly, a 4.6-bit quantization scheme is disclosed that produces a 4.6-bit QNN with similar accuracy to an 8-bit QNN, but a speed and computational efficiency that is similar to 4-bit QNNs.

Abstract: Quantization of a convolutional neural network (CNN) into a quantized neural network (QNN) reduces the computational resources required to operate the neural network, which is especially advantageous for operation of a neural network on resource-constrained devices. However, QNNs with low bit-widths suffer from significant losses in accuracies. Accordingly, approaches for quantization-aware training are disclosed that utilize component-by-component quantization during training to improve the accuracy of the resulting QNN. Component-by-component quantization may include filter-by-filter quantization, or preferably neuron-by-neuron quantization with some form of gradient forwarding.

Abstract: A bipolar morphological neural network may be generated by converting an initial neural network by replacing multiplication calculations in one or more convolutional layers with approximations that utilize maximum/minimum and/or addition/subtraction operations. The remaining part of the network may be trained after each convolutional layer is converted.

Abstract: No computationally efficient CPU-oriented algorithms of ternary and ternary-binary convolution and/or matrix multiplication are available. Accordingly, a microkernel is disclosed for high-performance matrix multiplication of binary, ternary, and ternary-binary matrices for central processing units (CPUs) with the Advanced Reduced Instruction Set Computer (RISC) Machine (ARM) v8 architecture.

Abstract: Approximate modeling of next combined result for stopping text-field recognition in a video stream. In an embodiment, text-recognition results are generated from frames in a video stream and combined into an accumulated text-recognition result. A distance between the accumulated text-recognition result and a next accumulated text-recognition result is estimated based on an approximate model of the next accumulated text-recognition result, and a determination is made of whether or not to stop processing based on this estimated distance. After processing is stopped, the final accumulated text-recognition result may be output.

Abstract: Artificial intelligence using convolutional neural network with Hough Transform. In an embodiment, a convolutional neural network (CNN) comprises convolution layers, a Hough Transform (HT) layer, and a Transposed Hough Transform (THT) layer, arranged such that at least one convolution layer precedes the HT layer, at least one convolution layer is between the HT and THT layers, and at least one convolution layer follows the THT layer. The HT layer converts its input from a first space into a second space, and the THT layer converts its input from the second space into the first space. The CNN may be applied to an input image to perform semantic image segmentation, so as to produce an output image representing a result of the semantic image segmentation.

Abstract: Image quality assessment for text recognition in images with projectively distorted text fields. A projective transformation is calculated from a restored rectangle, representing a restored text field, to a source quadrangle, representing a projectively distorted text field in a source image. An approximation of a curve of a minimal scaling coefficient level on a plane corresponding to the restored rectangle is constructed, based on calculations of a discriminant of the curve. When the approximation intersects a representation of the restored rectangle, a restoration of the source image is determined to have insufficient image quality for reliable text recognition. When the approximation does not intersect the representation of the restored rectangle, a minimal scaling coefficient is calculated at a point inside the restored rectangle, and a determination of whether or not the restoration of the source image has sufficient image quality is made based on the minimal scaling coefficient.

Abstract: Approximate modeling of next combined result for stopping text-field recognition in a video stream. In an embodiment, text-recognition results are generated from frames in a video stream and combined into an accumulated text-recognition result. A distance between the accumulated text-recognition result and a next accumulated text-recognition result is estimated based on an approximate model of the next accumulated text-recognition result, and a determination is made of whether or not to stop processing based on this estimated distance. After processing is stopped, the final accumulated text-recognition result may be output.

Abstract: A system for monitored tomographic reconstruction, comprising: an x-ray generator configure to generate x-ray beams for scanning an object; detectors configured to capture a plurality of projections for each scan; at least one hardware processor; and one or more software modules that, when executed by the at least one hardware processor, receive the plurality of projections from the detectors and as each of the plurality of projections is received, generate a partial reconstruction, and make a stopping decision with respect to whether or not another projection should be obtained based on a stopping problem and that defines when a reconstructed image quality is sufficient with respect to the expended cost as determined by a stopping rule.

Abstract: Artificial intelligence using convolutional neural network with Hough Transform. In an embodiment, a convolutional neural network (CNN) comprises convolution layers, a Hough Transform (HT) layer, and a Transposed Hough Transform (THT) layer, arranged such that at least one convolution layer precedes the HT layer, at least one convolution layer is between the HT and THT layers, and at least one convolution layer follows the THT layer. The HT layer converts its input from a first space into a second space, and the THT layer converts its input from the second space into the first space. The CNN may be applied to an input image to perform semantic image segmentation, so as to produce an output image representing a result of the semantic image segmentation.

Abstract: Real-time monitored computed tomography (CT) reconstruction for reducing a radiation does. During helical CT scanning of a target object, projections may be acquired in either a full mode which subjects the target object to a full radiation dose, or a reduced mode which subjects the target object to a reduced radiation dose (e.g., by reducing the number of projections acquired, reducing the exposure time, etc.). After a sector is acquired in the full mode, a slice of the target object that is influenced by that sector is identified, and a CT image of that slice is reconstructed using projections that have been previously acquired for that slice. When a stopping rule is satisfied based on this partial reconstruction, the full mode is switched to the reduced mode, and at least one subsequent sector is acquired in the reduced mode.

Abstract: A system for monitored tomographic reconstruction, comprising: an x-ray generator configure to generate x-ray beams for scanning an object; detectors configured to capture a plurality of projections for each scan; at least one hardware processor; and one or more software modules that, when executed by the at least one hardware processor, receive the plurality of projections from the detectors and as each of the plurality of projections is received, generate a partial reconstruction, and make a stopping decision with respect to whether or not another projection should be obtained based on a stopping problem and that defines when a reconstructed image quality is sufficient with respect to the expended cost as determined by a stopping rule.

Abstract: Advanced Hough-Based On-Device Document Localization. In an embodiment, lines are detected in an input image of a document. The lines are searched for candidate quadrilaterals. For at least a subset of the found candidate quadrilaterals, a contour score is calculated, and the candidate quadrilaterals are saved or discarded based on their contour scores. For each saved candidate quadrilateral, a contrast score is calculated. A final candidate quadrilateral is selected, based on the combined contour and contrast scores for the saved candidate quadrilaterals, to represent the borders of the document.

Abstract: Text recognition in a video stream using combined recognition results with per-character weighting. In an embodiment, for each frame in a video stream, a text-recognition result is obtained and a frame weight is calculated. The text-recognition results of the frames are combined by aligning character-recognition results and calculating a character weight for each character-recognition result. At each position in the alignment, the character-recognition results are accumulated based on the character weights and frame weights to produce an accumulated text-recognition result that represents a text field in the video stream.

Abstract: Reducing false detections in template-based classification of identity documents. In an embodiment, an iterative procedure is used to generate one or more hypotheses for the location of a document in image data and a type of document in the image data based on a plurality of predefined models representing a plurality of types of documents. The one or more hypotheses are filtered by rejecting any hypothesis that is not well-conditioned according to one or more criteria. When a best hypothesis that satisfies a threshold remains after filtering the one or more hypotheses, the document in the image data is analyzed, and, when no hypothesis that satisfies the threshold remains after filtering the one or more hypotheses, the image data is rejected.

Abstract: Unified framework for analysis and recognition of identity documents. In an embodiment, an image is received. A document is located in the image and an attempt is made to identify one or more of a plurality of templates that match the document. When template(s) that match the document are identified, for each of the template(s) and for each of one or more zones in the template, a sub-image of the zone is extracted from the image. For each extracted sub-image, one or more objects are extracted from the sub-image. For each extracted object, object recognition is performed. This may be done over one iteration (e.g., for a scanned image or photograph) or a plurality of iterations (e.g., for a video). Document recognition is performed based on the one or more templates and the results of the object recognition, and a final document-recognition result is output.

Abstract: Artificial intelligence using convolutional neural network with Hough Transform. In an embodiment, a convolutional neural network (CNN) comprises convolution layers, a Hough Transform (HT) layer, and a Transposed Hough Transform (THT) layer, arranged such that at least one convolution layer precedes the HT layer, at least one convolution layer is between the HT and THT layers, and at least one convolution layer follows the THT layer. The HT layer converts its input from a first space into a second space, and the THT layer converts its input from the second space into the first space. The CNN may be applied to an input image to perform semantic image segmentation, so as to produce an output image representing a result of the semantic image segmentation.

Abstract: A method for detecting security holograms on documents in a video stream is disclosed, including: searching for interest points and calculating descriptors in a frame; filtering of interest points in the previous frame so that only points located inside the quadrangle of the outer borders of the document remain; matching the descriptors of interest points of the current and previous frames; application of an algorithm for estimating the parameters of projective transformation between the frames; projective transformation of the quadrangle of the outer boundaries of the document from the previous frame to obtain the outer boundaries of the document in the current frame; document image normalization; calculating the color saturation and hue; updating the saturation and hue values; further considering the pixels of the normalized document image with brightness values not exceeding a preset threshold; filtration of the obtained image.