Fig. 2
From: Camouflaged object detection via context and texture-aware hierarchical interaction
The overall framework of the proposed CTHINet. The framework is divided into two branches: the texture encoder and the context encoder. We employ PVTv224 as the context encoder and construct the texture encoder using several consecutive convolutional blocks. The output features of the context encoder are delivered to the MFAM to refine them individually, aiming to capture rich multi-scale contextual information in each surrounding layer. Subsequently, the texture and context features are fed through the HMIM module to aggregate features. The network follows a coarse-to-fine structure to enhance the camouflaged features progressively. (The images are sourced from the publicly available COD dataset: CAMO (https://sites.google.com/view/ltnghia/research/camo).
