Abstract: Provided are machine learning systems and models featuring resolution-flexible multi-axis attention blocks. In particular, the present disclosure provides example multi-axis MLP based architectures (example implementations of which can be generally referred to as MAXIM) that can serve as an efficient and flexible general-purpose vision backbone for image processing tasks. In some implementations, MAXIM can use a UNet-shaped hierarchical structure and supports long-range interactions enabled by spatially-gated MLPs. Specifically, some example implementations of MAXIM can contain two MLP-based building blocks: a multi-axis gated MLP that allows for efficient and scalable spatial mixing of local and global visual cues, and a cross-gating block, an alternative to cross-attention, which accounts for cross-feature mutual conditioning.

Abstract: Provided is an efficient and scalable attention model that can be referred to as multi-axis attention. Example implementations can include two aspects: blocked local and dilated global attention. These design choices allow global-local spatial interactions on arbitrary input resolutions with only linear complexity. The present disclosure also presents a new architectural element by effectively blending the proposed multi-axis attention model with convolutions. In addition, the present disclosure proposes a simple hierarchical vision backbone, example implementations of which can be referred to as MaxViT, by simply repeating the basic building block over multiple stages. Notably, MaxViT is able to “see” globally throughout the entire network, even in earlier, high-resolution stages.

Abstract: The technology provides an image resizer that is jointly trainable with neural network classification (recognition) models, and is designed to improve classification performance. Systems and method include applying an input image to a baseline resizer to obtain a default resized image, and applying the input image to a plurality of filters. Each respective filter in the plurality is configured to perform sub-band filtering on the input image to obtain a sub-band filtered result. This includes applying the sub-band filtered result to the baseline resizer to obtain a respective resized result, and also includes applying to the respective resized result a scaling parameter, a bias parameter, and a nonlinear function to obtain a respective filtered image. The process then combines the default resized image and the respective filtered images to generate a combined resized image.

Abstract: A method of measuring a banding artefact in an image includes generating a gradient profile from the image, where the gradient profile includes respective gradient magnitudes of pixels of the image; generating, using the gradient profile, a candidate banding pixel (CBP) map, where each location of the CBP map is such that a gradient magnitude of the gradient profile of a corresponding pixel of the image being greater than a first threshold and smaller than a second threshold; generating, using the CBP map, a banding edge map (BEM), where the BEM includes connected banding edges of the image; generating, using the BEM, a banding visibility map (BVM), where the BVM includes a respective banding metric for at least some pixels of the image; and generating a banding index of the image using the BVM.

Abstract: A method of measuring a banding artefact in an image includes generating a gradient profile from the image, where the gradient profile includes respective gradient magnitudes of pixels of the image; generating, using the gradient profile, a candidate banding pixel (CBP) map, where each location of the CBP map is such that a gradient magnitude of the gradient profile of a corresponding pixel of the image being greater than a first threshold and smaller than a second threshold; generating, using the CBP map, a banding edge map (BEM), where the BEM includes connected banding edges of the image; generating, using the BEM, a banding visibility map (BVM), where the BVM includes a respective banding metric for at least some pixels of the image; and generating a banding index of the image using the BVM.