Convolutional Neural Networks on Programmable 2D Image Processors

Abstract

A method is described that includes executing a convolutional neural network layer on an image processor with an array of execution channels and a two-dimensional shift register. Two-dimensional shift registers provide locally corresponding register spaces for execution channels. Execution of the convolutional neural network includes loading a plane of image data of a three-dimensional image data block into a two-dimensional shift register. Executing the convolutional neural network also includes performing a two-dimensional convolution of the plane of the image data with the array of coefficient values ​​by sequentially: multiplying corresponding pixels and coefficient values ​​simultaneously within an execution pass to produce an array of partial products; simultaneous in-channel summation of partial products and corresponding cumulative values ​​of partial products held in two-dimensional registers of different templates within the image data; and performing in-channel two-dimensional The alignment of the convolved values.

Description

technical field [0001] This description relates generally to image processing, and more specifically, to a convolutional neural network on a two-dimensional image processor. Background technique [0002] Image processing typically involves manipulating pixel values ​​organized into arrays. Here, the spatially organized two-dimensional array captures the two-dimensional nature of the image (additional dimensions may include time (eg, a series of two-dimensional images) and data type (eg, color)). Typically, the arrayed pixel values ​​are provided by a camera that produces a still image or a series of frames to capture a moving image. Traditional image processors typically fall into one of two extremes. [0003] The first extreme performs image processing tasks by executing software programs on a general purpose processor or a general purpose-like processor (eg, a general purpose processor with vector instruction enhancement). While the first extreme typically provides a hi...

AI Technical Summary

Problems solved by technology

However, the use of custom-designed fixed-function circuits typically results in a limited set of tasks that the processor can perform

Thus, the second extreme (associated with the first) lacks a broadly general programming environment

Images