System for providing enhanced vision transformer blocks for computer vision

Abstract

A system for providing an enhanced vision transformer block for mobile vision transformers to perform computer vision tasks, such as image classification, segmentation, and objected detection is disclosed. A local representation block of the block applies a depthwise-separable convolutional layer to vectors of an input image to facilitate creation of local representation outputs associated with the image. The local representation output is fed into a global representation block, which unfolds the local representation outputs, applies vision transformers, and folds the result to generate a global representation output associated with the image. The global representation output is fed to a fusion block, which concatenates the local representations with the global representations, applies a point-wise convolution to the concatenation to generate a fusion block output, and fuses input features of the image with the fusion block out to generate an output to facilitate performance of a computer vision tasks.