Systems and methods are described herein that use passive metasurfaces for optical domain image processing, reducing power consumption, latency, and/or computational load in computer vision. As an example, an optical detection device may include an optical assembly that generates multiple copies of an image, which are processed by an array of metasurface filters. Each metasurface filter applies a predefined pattern to a specific image copy. The filtered image copies are used to identify features of interest with reduced computational demands. For example, a first filter may target a first feature, while a second filter targets a second feature. The intensity of the optical radiation output by each filter is used to determine a confidence level that each respective feature is present in the image.