Low-light Image Enhancement (LLIE) aims to correct insufficient illumination and improve overall visual quality. However, existing methods often struggle to balance global illumination adjustment with local contrast enhancement, attempting to address coupled degradations within a unified architecture, leading to suboptimal trade-offs. To address this, we propose TG-SFCnet, a novel task-guided Spatial-Frequency Collaborative Network with a two-stage cascaded architecture. TG-SFCnet strategically directs each sub-task to its most suitable processing domain: global illumination recovery is guided to the spatial domain, while local detail refinement is directed to the frequency domain. In the first stage, a Gated Multi-scale Aggregation Network (GMANet) equipped with a Pixel Gated Convolution (PGC) module restores global illumination in the spatial domain. In the second stage, a Multi-Branch Fusion Attention (MBFA) mechanism is developed in the frequency domain to suppress noise and preserve texture details effectively. Experimental results show that our framework consistently outperforms state-of-the-art methods.
The Multi-Focus Image Fusion (MFIF) technique integrates various source images into a single clearer image, with the identification of clear points being crucial for effective fusion. Traditional edge-preserving filters struggle with scale-dependent edge identification and necessitate guide images. This research introduces a ResNet-based image super-resolution technique paired with Non-Subsampled Shearlet Transform (NSST) Coefficient Extraction and Energy of Gradient (EOG) focus measurement. It employs a small area elimination process and a rolling guidance filter to enhance the initial decision map. The fusion of weighted source images and decision maps yields superior results, with performance metrics like MI of 9.3670, QY of 0.9770, and a runtime of 60.23 s, outperforming existing models such as SR, BF, and CNN variants. This framework significantly enhances resolution while preserving structural details in multi-focus image fusion.
To address low resolution, blurred details, and colour distortion in fashion illustration images for digital education, a Computational Intelligence (CI)-based image processing method is proposed. An eight-layer Deep Convolutional Neural Network (DCNN) extracts and enhances image features using the Leaky ReLU activation function, performing feature extraction, nonlinear mapping, enhancement, and reconstruction. DCNN parameters are optimized via a hybrid Artificial Bee Colony-Particle Swarm Optimization strategy, with PSNR as the fitness function and Adam gradient optimization for accurate mapping from low- to high-resolution images. Experiments show structural similarity above 96%, PSNR up to 34.56 dB, and colour restoration of 99.34%, significantly improving edge clarity, texture detail, and brightness uniformity. The method restores line continuity, colour transitions, and fabric textures, enhancing both visual quality and the effectiveness of digital fashion education.
A multi-feature adaptive correlation filter (MFACF-3D) was proposed in the study for target tracking in 3D image sequences, aiming to address the drift problems caused by occlusion, scale changes and complex scenes. It collaboratively integrates deep guidance filtering and the eight-parameter motion model to conduct explicit 3D geometric reasoning. By fusing HOG, color and depth features, its accuracy rate reached 91.5%, the success rate was 98% (with a 300-pixel threshold), and the success rate in the presence of occlusion was 90% - more than 15% higher than traditional methods. All improvements were statistically significant (p < 0.01).
Infrared and visible image fusion is crucial for multimodal perception, but high-performance models incur excessive computational costs for resource-constrained edge devices. This work introduces TL2Fusion, a unified compression-to-deployment framework that transforms the SeAFusion network into a hardware-efficient implementation. First, a dual-stream-adapted Tucker decomposition achieves substantial model compression. Second, a structured pruning strategy leverages hierarchical dependency analysis and L2-norm-based filter sensitivity to intelligently remove redundant convolutional filters. Finally, a targeted retraining phase restores model accuracy. Evaluations on TNO, FMB, and MSRS datasets demonstrate competitive fusion quality and superior performance in downstream segmentation tasks. Furthermore, by optimizing the GRDB module for FPGA deployment, TL2Fusion reduces memory footprint by 71.3-73.2% and improves power efficiency by 64.2% with minimal computational overhead. This research provides a validated pipeline for deploying real-time, energy-efficient multimodal perception systems at the edge.
This study addresses the accuracy decline of traditional digital holographic measurement under high-noise conditions. A novel phase reconstruction algorithm is proposed to prevent the premature loss of key information. By representing the light field as a complex vector, complex-domain calculation is realized with embedded complex amplitude information, which improves noise immunity and phase calculation reliability. Combined with a one- dimensional Fourier transform-based time-domain phase extraction algorithm, spatial propagation of phase errors is restrained. The effect of decorrelation error is analyzed, and optimal conditions are derived to further enhance denoising performance. Experimental results confirm that the proposed algorithm achieves superior denoising ability and measurement reliability over conventional methods.
A 3D face reconstruction approach using advanced deep learning techniques is implemented to create avatars. At first, 2D face images are obtained from standard repositories. These images are given to the proposed deep learning technique called Adaptive Multi-Scale Dense 3D Convolutional Generative Adversarial Network with Axial Spatial Attention (AMD-3DCGAN-SA) model for 3D face reconstruction. The 3DCNN model effectively reconstructs facial features besides 3D face imagery. Global spatial features in face images are precisely analyzed by the axial spatial attention mechanism to preserve all the important facial details. Furthermore, an optimization algorithm known as Fitness Normalization-based Willow Catkin Optimization (FNWCO) is used for fine-tuning model parameters for accurate 3D facial modelling. Overall, the effectiveness of the developed approach was evaluated against traditional face reconstruction methods to ensure its efficiency.
Hierarchical image detection in high-resolution remote sensing utilizes a structured approach to address the challenges posed by varying target sizes and proportions. The process begins with segmenting the target foreground using a hierarchical Gaussian mixture model. This is followed by employing texture constraints between pixels within the target foreground, aided by least squares support vector machines (LSSVM) for initial target detection. Next, attribute constraints between neighbouring objects help correct initial detection errors. Finally, spatial distribution constraints and category characteristic distribution constraints refine the results, addressing mislabelling at the hierarchical level. Experimental results demonstrate that this method effectively classifies ground objects and detects vehicle targets in high-resolution remote sensing images, enhancing detection accuracy amidst spatial and semantic complexities.
