2 June 2026 in SSRN Electronic Journal
Low-light image enhancement aims to recover well-exposed images from degraded observations. However, existing curve-based zero-reference methods are often limited by shallow architectures with restricted receptive fields and by uniform enhancement strategies that tend to over-enhance already bright regions. To address these issues, we propose HDAR-Net, a Hierarchical Dual-Attention network with Regional protection, which integrates three complementary innovations. First, a hierarchical U-Net encoder-decoder expands the receptive field from 7 × 7 to 127 × 127 pixels, enabling more effective global context modeling. Second, a dual-branch CBAM attention mechanism performs sequential channel and spatial attention, allowing content-adaptive feature selection. Third, a content-aware regional protection module with learnable brightness estimation adaptively regulates enhancement strength, thereby suppressing over-saturation artifacts in bright areas. Extensive experiments on the LOL and SICE benchmarks show that HDAR-Net achieves state-of-the-art performance, reaching PSNR values of 20.16 dB and 18.02 dB, respectively, while remaining computationally efficient with only 2.29 million parameters and an inference speed of 125 FPS, demonstrating its practicality for real-world deployment.
11 March 2026 in SSRN Electronic Journal
In open-world dialogue systems, new intent discovery (NID) aims to identify previously unseen user intents from raw utterances and expand the supported intent set. This process is traditionally time-consuming and often requires domain experts to explore data and define intent labels, which substantially increases the cost. To address this challenge, we propose SCNID, a semi-supervised framework that integrates contrastive representation learning with LLM-driven pairwise-constraint-guided clustering fusion. Specifically, SCNID learns discriminative intent representations via semi-supervised contrastive learning, and then performs clustering fusion by combining a Gaussian Mixture Model (GMM) with K-means++ to improve robustness and stability. Moreover, we leverage a large language model (LLM) to generate Must-Link and Cannot-Link constraints, providing structured semantic guidance for clustering. Extensive experiments on four benchmarks---three intent discovery datasets (Clinc150, Banking77, and StackOverflow20) and an additional cross-domain topic dataset (DBPedia14)---demonstrate that SCNID achieves strong and stable performance under diverse data distributions, validating the effectiveness of the proposed approach.