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: 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: Almost-indirect convolution in quantized neural networks (QNNs). In an embodiment, an indirection buffer is allocated and initialized with pointers to block of values in an input tensor representing an input image. In one or more layers of a QNN, matrix multiplication of a left matrix by a right matrix is performed over a plurality of iterations. In each iteration, a block of one or more rows of the input tensor is packed into a first cache using the indirection buffer, and a convolution kernel is applied to the packed block to produce an output. The outputs are accumulated as an output tensor representing an output image.

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.