Camouflaged object detection (COD) aims to identify objects that exhibit high visual similarity to their surroundings. Existing frequency- and edge-assisted methods improve feature discrimination through spectral and boundary cues, yet these cues do not consistently benefit both region completeness and boundary localization when introduced independently at different decoding stages, particularly in scenes containing high-frequency background clutter. To address this issue, we propose EGBDNet, an explicit structural prior learning framework that transforms heterogeneous structural cues into a reusable prior representation for progressive feature learning. Specifically, CEPM constructs the structural prior by integrating multi-scale frequency responses with semantic information under edge supervision. EBDC exploits this prior to organize body–edge representation learning, while CAGM regulates the residual enhancement of the structural prior according to decoder states. Experiments on CAMO, COD10K, and NC4K demonstrate competitive performance against representative COD methods.
This paper studies Conv-LISTA, a LISTA-inspired convolutional deep-unfolding architecture for reconstruction from irregularly missing samples on discrete one-dimensional (1D) and two-dimensional (2D) grids. Each stage combines learned observation/iterate convolutions, synthesis and soft-thresholding with an exact hard data-consistency projection at observed samples. The 1D and 2D instantiations share the same stage topology but use dimension-specific filters and independently trained weights.A finite-depth perturbation recursion is derived to separate conditional stability from global convergence. Soft-thresholding remains the proximal map of an ℓ1 penalty, but the complete learned convolutional block is not generally a convex proximal map. Numerical stage-norm diagnostics further show that none of the learned stages satisfies a conservative mask-uniform contraction certificate; consequently, global convergence is not claimed, while failure of the sufficient certificate is not interpreted as instability.Experiments use synthetic 1D signals and MNIST under random and structured missingness. At the central random-missingness condition (ρ=0.5, σ=0.01), Conv-LISTA obtains 34.91 dB missing-only SNR in 1D and 10.26 dB in 2D, but stronger baselines exist in both tasks. Structured contiguous/rectangular missing blocks cause substantial degradation, exposing a clear limitation of the local convolutional prior. Conv-LISTA is nevertheless highly parameter-efficient: the selected models contain 6980 parameters in 1D and 53,670 in 2D. The study therefore characterizes Conv-LISTA as an interpretable and compact fixed-depth reconstruction model with exact sample consistency, while quantifying the conditions under which its local inductive bias and unconstrained stage weights limit robustness and theoretical guarantees.
