Aiming at the shortcomings of existing low-light image enhancement methods, such as poor illumination adaptability and difficulty in balancing global structure and local details, this paper proposes a light-prior driven low-light enhancement network (LPD-Net). First, an illumination prior module (IPM) is constructed based on Retinex theory, which generates illumination-normalized features by introducing homomorphic filtering priors to achieve global correction of non-uniform illumination. Second, a Mamba enhanced module (MEM) based on an encoder-decoder architecture is designed, employing a hierarchical modeling strategy and constructing a multi-scale pathway consisting of local window Mamba and global Mamba to achieve local detail enhancement and global brightness restoration. On this basis, a dual-stream fusion color modulation (DFCM) module is introduced, which captures horizontal and vertical spatial dependencies through bidirectional convolutional blocks and integrates global-local features using a lightweight attention mechanism. Simultaneously, channel gating and color modulation mechanisms are applied to adaptively restore color information and fine image details. Finally, total variation regularization is introduced to constrain the illumination map, preventing structural inconsistency caused by over-enhancement. Experimental results show that the proposed method achieves 24.370 dB PSNR and 0.853 SSIM on the LOLv1 dataset, significantly outperforming existing methods. The method demonstrates notable advantages in illumination correction, noise suppression, and detail preservation, while also exhibiting strong generalization capability for real-world low-light scenarios.

In crowd counting, traditional methods often struggle to balance preserving local details and capturing global context due to the high variability in crowd density and occlusion levels within scenes, leading to performance degradation in highly congested or severely perspective-distorted images. To address this issue, this paper proposes a dense-scene crowd counting method based on deep spectral adaptive modulation (DSAMNet), aiming to simultaneously enhance the spatial resolution of density map estimation and global counting consistency. Specifically, the method first employs a depth-aware module to jointly extract multi-scale texture features and coarse depth estimates, utilizing predicted depth to guide perspective correction and achieve geometrically consistent feature sampling. Subsequently, a spatially adaptive high-frequency encoding module is introduced, which dynamically modulates coordinate frequencies based on local depth during the encoding process, thereby enhancing the positional representation capacity in geometrically sensitive regions and improving the network's adaptability to scale and structural variations. Finally, an implicit density decoding module integrates visual and geometric representations through a cross-domain attention mechanism and performs point-wise density regression via a hierarchically conditioned multi-layer perceptron, achieving high-quality reconstruction of continuous density fields. Experimental results demonstrate that the proposed method achieves superior performance on several mainstream crowd counting datasets, exhibiting strong robustness and generalization capability in complex perspective scenarios.

The quality of super-resolution images is not only affected by the reconstruction algorithm, but also there are some differences in the quality degradation levels of the reconstructed images under different upscaling-factors. However, the existing no-reference super-resolution image quality assessment (NR-SRIQA) methods mainly focus on the visual features of super-resolution images, ignoring the available upscaling-factor information. An upscaling-factor aware contrastive learning (UFACL) method is proposed. The network structure is divided into a upscaling-factor recognition branch and a quality score branch. The upscaling-factor recognition branch starts from the dataset, and takes the super-resolution images of different upscaling-factors as positive and negative samples of each other. Contrastive learning is introduced to complete the classification task, so as to improve the expression ability of effective features. In the quality score branch, a frequency domain attention module (FDAM) is designed, which considers both global information and channel information. At the same time, this branch uses inverted residuals blocks (IRB) to reduce the calculation amount of the model, which ensures the accuracy of quality score prediction and improves the training efficiency of the model in the training process. Experimental results show that the proposed UFACL can achieve better consistency with subjective perceived quality.