Progressive exploration detection method for camouflage small target

Abstract

This invention discloses a progressive exploratory detection method for camouflaged small targets, relating to the field of camouflaged target detection. The method includes: inputting an RGB image into an encoder; expanding the receptive field using an ASPP module to generate hierarchical features; aligning and fusing the hierarchical features using a multi-scale feature enhancement module (MFEM) to re-inject deep semantics into the high-resolution stream; generating a coarse localization mask based on a context-guided attention mechanism (CGA) and fused features; iteratively refining the coarse localization mask using a time-series regression network; updating the mask through a gated regression mechanism; and combining expansion and contraction operators to capture weak boundaries and suppress overgrowth; performing end-to-end training using a composite loss function; and employing a progressive localization mechanism from coarse to fine during the inference phase to output the final segmentation result. This invention solves the problem of existing technologies struggling to handle small targets and blurred boundaries.

Images