The growing interest in image dehazing has led to significant advancements in deep learning-based single image dehazing approaches, achieving remarkable performance in recent studies. However, many existing methods fail to meet the efficiency demands of practical applications. In this paper, we introduce WaveDH, a novel and compact ConvNet designed to address this efficiency gap. Our WaveDH leverages wavelet sub-bands for guided up-and-downsampling and frequency-aware feature refinement. The key idea lies in utilizing wavelet decomposition to extract low- and high-frequency components from feature levels, allowing for faster processing while upholding high-quality reconstruction. The downsampling block employs a novel squeeze-and-attention scheme to optimize the feature downsampling process in a structurally compact manner, preserving discriminative features while discarding noise components. In our upsampling block, we introduce a dual-upsample and fusion mechanism to enhance high-frequency component awareness, aiding in the reconstruction of fine details. Departing from conventional dehazing methods that treat frequency components equally, our feature refinement block adopts a frequency-aware, coarse-to-fine approach to process features more effectively. By refining details in a maximum 8 × downsampled feature space, WaveDH strikes a favorable trade-off between accuracy and computational cost. Extensive experiments demonstrate that WaveDH outperforms many state-of-the-art methods across several benchmarks. Notably, on the SOTS-indoor dataset, WaveDH achieves a peak signal-to-noise ratio (PSNR) of 39.35 dB and a structural similarity index (SSIM) of 0.995, surpassing existing high-complexity models while requiring significantly reduced computational costs. Our code is available at https://github.com/smin-hwang/WaveDH .
This study introduces a computationally efficient and robust pipeline for rail defect detection that strategically combines shallow feature extraction with selective deep segmentation to achieve high precision with minimal computational overhead. We curated a dataset of 356 cropped rail surface images (132 Normal, 101 Joint, 123 Defective) from multiple public and private sources under varying real-world conditions, permanently partitioned into 249 training (70
Future trajectories of neighboring traffic agents heavily influence the path planning and decision-making of autonomous vehicles. While traditional trajectory forecasting focuses on snapshot-based predictions that treat scenarios independently, real-world deployment demands continuous, real-time data stream processing with low latency and temporal consistency. To address these requirements, we introduce a lightweight, streaming-based trajectory forecasting framework that propagates temporal context across successive timesteps. By utilizing the predicted endpoints of previous forecasts as structural anchors, our approach efficiently guides a scene encoder to extract highly targeted context encodings without requiring multi-stage refinement or segment-wise decoding. To ensure deterministic execution bounds on real-world robotics hardware, we enforce static token boundaries via proximity-based sampling to satisfy edge-compilation constraints. Evaluated on the Argoverse 2 multi-agent and single-agent benchmarks, our method achieves state-of-the-art streaming accuracy while requiring fewer computational resources than snapshot baselines. Furthermore, hardware deployment tests on an embedded NVIDIA Jetson Orin Nano platform demonstrate that our model maintains an inference frequency exceeding 10 Hz while preserving a highly favorable latency-accuracy trade-off, proving its viability for deterministic, low-power, real-time execution in dense, continuous traffic environments.
During welding, it is necessary to analyze the welding penetration state by evaluating the morphological characteristics of the welding pool to promptly identify welding deficiencies and implement corresponding adjustments. This study establishes a full-position pipeline welding penetration prediction model based on a backpropagation neural network (BPNN) by integrating automation and computer image processing technologies and combining weld-pool information with full-position welding parameters. The study systematically analyzes the force conditions of a weld pool in different zones during full-position pipeline welding and summarizes the morphological characteristics of the weld pool in full-position welding scenarios. A visual sensor was employed to monitor the weld pool morphology, enabling the observation of pool shape characteristics at different spatial positions. A frontal geometric model of the weld pool was constructed, and a highly robust image-processing algorithm was designed to effectively extract the geometric characteristic parameters of the pool, despite arc light interference and spatter noise. An position-aware BPNN was used to develop a penetration prediction model for root welding during fully positioned pipeline welding. The model used the extracted pool characteristic parameters and spatial position data of the welding robot as input variables, and the penetration state as the output. Experimental validation demonstrates that the model achieves a prediction accuracy of 87
SIGMAR-3D is a training-free framework for robust pairwise 3D point-cloud registration. It combines a handcrafted local descriptor, SIGMA, with a scene-adaptive coarse-to-fine registration stage, R-3D. SIGMA builds a repeatable local reference frame at each keypoint and aggregates spatial-angular geometric cues within a spherical support, producing a compact rotation-normalized descriptor without any learning. R-3D forms tentative correspondences by descriptor matching, estimates a coarse pose using robust hypothesis verification, and refines the alignment on a voxel pyramid with generalized ICP and robust losses. Key support radii, matching gates, and refinement thresholds are tied to local spacing statistics and overlap-related heuristic signals, enabling a single configuration to transfer across indoor RGB-D and outdoor LiDAR data. Monotone decrease of the trimmed-RMSE refinement objective is enforced through an acceptance rule that accepts an update only when that objective decreases. On standard benchmarks, SIGMAR-3D achieves 84.3
