Answers are grounded in this paper and cite their sources.
Camouflaged object detection (COD) aims to segment objects that are seamlessly embedded in their surroundings. However, existing models still struggle with extreme camouflage, where local textures are nearly indistinguishable from the background. Motivated by a cognitive analogy, in which observers reduce perceptual ambiguity by imagining a plausible target appearance before verification, we propose a Cognitive Visual Imagination (CVI) framework. CVI is divided into two stages: (1) Mental Template Generation, which uses the generative prior of a frozen pre-trained image editing foundation model to create a cognitive template that highlights the target while suppressing background distractions; and (2) CVINet, a dual-stream segmentation network that combines the original image with the generated template. To bridge low-level visual cues and high-level cognitive priors, CVINet uses a weight-sharing encoder to extract multi-scale features and introduces two modules inspired by complementary perceptual processes: (i) the cognitive semantic injection (CSI) module, which adaptively injects high-level semantic cues based on cross-stream feature discrepancies, motivated by contrastive perception, and (ii) the cognitive structure refinement (CSR) module, which uses high-frequency priors to bolster structural details and refine object boundaries, analogous to boundary confirmation. Extensive experiments with four widely used COD benchmark datasets show competitive performance. By operationalizing a cognitively inspired “imagine-to-verify” workflow, this study provides an effective and practical approach to camouflaged object detection. The source code will be made publicly available upon acceptance.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: