Residual U-Net beats prior cloud segmentation on all four metrics
The k=4 model with deep supervision reaches F-measure 0.93 on SWINySEG in under 17,500 iterations.
· “UCloudNet: A Residual U-Net with Deep Supervision for Cloud Image Segmentation”
eess.IV
Image and Video Processing
Theory, algorithms, and architectures for the formation, capture, processing, communication, analysis, and display of images, video, and multidimensional signals in a wide variety of applications. Topics of interest include: mathematical, statistical, and perceptual image and video modeling and representation; linear and nonlinear filtering, de-blurring, enhancement, restoration, and reconstruction from degraded, low-resolution or tomographic data; lossless and lossy compression and coding; segmentation, alignment, and recognition; image rendering, visualization, and printing; computational imaging, including ultrasound, tomographic and magnetic resonance imaging; and image and video analysis, synthesis, storage, search and retrieval.
sort pith recommended most recent
2.5D super-resolution lifts CT defect detection at 2D cost
Feeding seven CT slices into a super-resolution net beats 2D on defect detection while adding under 3% memory.
Stacked-hourglass landmarks beat traditional models on palsy faces
On 87 facial palsy images, only the deep network localised mouth landmarks accurately enough for 3D modelling.
LED-only microscope maps 3D refractive index of whole worms
A nonlinear solver turns 120 intensity images into quantitative 3D structure at 250 nm resolution.
Content preservation splits image-to-image tasks into three buckets
A review of ~70 models and a new vehicle Sim2Real benchmark helps pick translators when source detail must survive.
· “Unpaired Image-to-Image Translation with Content Preserving Perspective: A Review”
Review charts deep learning for power line inspection
A structured look at component detection and fault diagnosis, and the data gaps still blocking automation.
· “Deep Learning in Automated Power Line Inspection: A Review”
0.33M-parameter U-Net hits 95.3% on cloud segmentation
Dynamic multi-scale and per-image classifier weights keep accuracy near far larger models on SWINySEG.
· “DDUNet: Dual Dynamic U-Net for Highly-Efficient Cloud Segmentation”
One 3D model learns MRI, CT, and PET without labels
Pretrained on ~99,000 unlabeled scans, 3DINO-ViT transfers to unseen organs and modalities.
· “A generalizable 3D framework and model for self-supervised learning in medical imaging”
Sub-meter radar land-cover benchmark opens all-weather mapping
SAR sees through clouds; this dataset lets models map land cover when optical imagery fails.
Stable Diffusion becomes a zero-shot lossy image compressor
First full DiffC implementation encodes in under 10 seconds and rivals purpose-built codecs at ultra-low bitrates.
Small diffusion priors beat big ones on medical scans
A 5.3-million-parameter model nearly matches attention U-Nets in-distribution and wins on new anatomies.
· “Bigger Isn't Always Better: Towards a General Prior for Medical Image Reconstruction”
Super-resolution models sharpen low-bit color recovery
Pre-trained SR encoders act as frozen feature extractors, lifting PSNR and SSIM beyond current bit-depth methods.
· “Bit-depth color recovery via off-the-shelf super-resolution models”
Bootstrapping a 4× tokenizer yields sharp 16× video latents
Grow the model in stages instead of training from scratch, and 16× latents cut diffusion cost 2.5× without quality loss.
· “Progressive Growing of Video Tokenizers for Temporally Compact Latent Spaces”
SplineFormer drives a robot to cannulate an artery 50% of the time
Predicting the guidewire as a smooth B-spline lets a robot navigate fully autonomously to the brachiocephalic artery.
· “SplineFormer: An Explainable Transformer-Based Approach for Autonomous Endovascular Navigation”
Graph explainer beats Grad-CAM and SHAP on breast tissue maps
Multiscale graph attention plus MIL and gradient fusion aligns AI heatmaps with pathologist-marked tumour regions.
Disease-aware memory lifts X-ray report clinical F1 by 19.5%
A two-memory design retrieves disease regions and past reports to sharpen LLM-written radiology text.
· “Activating Associative Disease-Aware Vision Token Memory for LLM-Based X-ray Report Generation”
3D vessel maps from sparse DSA images in 5 minutes
A 4D Gaussian-splatting model cuts reconstruction time from hours to minutes while preserving vessel detail.
Real-data benchmark puts 12 learned CT methods on equal footing
Five standardized tasks on the same measured scans, with open code, let any new algorithm be tested against the same baselines.
· “Benchmarking learned algorithms for computed tomography image reconstruction tasks”
Context modules lift traffic sign mAP to 92.1% on TT100K
Local and global context replace YOLOv7's neck blocks, raising accuracy while cutting parameters by 2.7M.
· “YOLO-CCA: A Context-Based Approach for Traffic Sign Detection”
Two cooperating networks lift underwater image PSNR to 24.54 dB
CoRe-Net restores underwater images with 7.2M parameters, about nine times fewer than the prior best.
· “Blind Underwater Image Restoration using Co-Operational Regressor Networks”
Vessel mesh generation drops from 2 hours to 30 seconds
A fixed graph template deformed by a GCN yields anatomically consistent hexahedral meshes for coronary and cerebral vessels.
Whole-slide model outdoes ROI models without finetuning
Pretrained on 335K slides and aligned with captions and reports, TITAN transfers to rare-cancer retrieval and zero-shot diagnosis.
Fixed points turn restorers into plug-and-play priors
Pair a pretrained restorer with its training degradation and an explicit prior appears.
· “FiRe: Fixed-points of Restoration Priors for Solving Inverse Problems”
Joint training sharpens motion estimates from undersampled MRI
Learning sampling, reconstruction, and registration in one network beats training them separately on cardiac and aorta scans.
· “Deep End-to-end Adaptive k-Space Sampling, Reconstruction, and Registration for Dynamic MRI”
Deformable registration maps 3D brain territories onto 2D DSA
Affine alone fails; B-spline registration reaches median SSIM 0.81 across 2,247 DSAs
Token-pruning video codec beats H.264+LDPC over noisy wireless links
Learned multi-scale spatial-temporal transformer trims low-importance tokens and holds PSNR above 30 dB at low bandwidth.
· “A Multi-Scale Spatial-Temporal Network for Wireless Video Transmission”
Thick clinical MRI slices sharpen to 1 mm without training
Multimodal scans are reconstructed from ~6.5 mm slices and automated brain analysis improves.
Cell graphs top 97% accuracy in colorectal cancer grading
Representing tissue as a network of nuclei captures the full micro-architecture that patch-based methods miss.
· “CGC-Net: Cell Graph Convolutional Network for Grading of Colorectal Cancer Histology Images”
Image-scale unbalanced transport drops from N^2 to N variables
A flux-field reformulation plus a parallel proximal solver puts unbalanced transport into large-scale inverse imaging.
· “Parallel Unbalanced Optimal Transport Regularization for Large Scale Imaging Problems”
SpaNet separates overlapping nuclei at 62.39% AJI
A proposal-free network clusters per-pixel spatial codes to beat previous best nuclear segmentation on a public multi-organ dataset.
· “Nuclear Instance Segmentation using a Proposal-Free Spatially Aware Deep Learning Framework”
Fusing per-scene re-exposed images beats two-image HDR fusion
Single-shot dual-ISO capture gains automatic scene-aware exposure compensation, winning where two-image fusion fails.
· “An Image Fusion Scheme for Single-Shot High Dynamic Range Imaging with Spatially Varying Exposures”
Monochrome camera sorts red, green, and blue parts with colored light
Matching filter and LED colors lets a monochrome camera sort red, green, and blue blocks.
· “Conveyor Line Color Object Sorting using A Monochrome Camera, Colored Light and RGB Filters”
Task decides which eye-AI model wins: DINOv2 or RETFound
Which model wins depends on the task: generalist for eye lesions, specialist for heart and stroke risk.
Four-network ensemble predicts rice yields within 341 kg/Ha
Fusing radar, optical, and weather data cuts the error to 5–6% of typical Mekong Delta rice yields.
· “Multi-modal Data Fusion and Deep Ensemble Learning for Accurate Crop Yield Prediction”
DPP band grouping proposed to sharpen hyperspectral reconstruction
Selecting diverse correlated bands with spectral-angle cleanup is meant to feed any ML model.
· “Leveraging band diversity for feature selection in EO data”
30x smaller model matches deep denoisers
A denoising-trained network transfers to super-resolution and MRI with only two tuned hyperparameters.
· “DEALing with Image Reconstruction: Deep Attentive Least Squares”
Fake MRIs in training cut tumor-segmentation Dice by 0.15
Raising GAN-made T1-Ce scans from 33% to 83% of the training set sinks overlap and recall on real brain MRIs.
· “Synthetic Poisoning Attacks: The Impact of Poisoned MRI Image on U-Net Brain Tumor Segmentation”
Uncertainty, not location, sets Mamba's scan order
Sorting pixels by channel standard deviation focuses Mamba on boundaries and foreground, lifting Dice on three medical datasets.
· “UD-Mamba: A pixel-level uncertainty-driven Mamba model for medical image segmentation”
Layer-separated X-rays synthesize adjustable joint-space images
Splitting bone from soft tissue yields synthetic X-rays with exact joint-space labels for training rheumatoid arthritis models.
· “Layer Separation: Adjustable Joint Space Width Images Synthesis in Conventional Radiography”
GANs get infinite semantic augmentation at near-zero cost
By perturbing deep features along estimated covariance directions, ASA improves few-shot synthesis without touching image pixels.
· “Adversarial Semantic Augmentation for Training Generative Adversarial Networks under Limited Data”
River connectivity images classified at 90% on unseen sites
Temporal luma averaging plus vision transformers raise accuracy from 75% to 90% on unseen stream sites.
Radar plus optical beats either sensor on a Cerrado test set
Vision-transformer baseline hits 57.60% macro F1 over 30,322 patches, showing how hard minority classes remain.
Automated PASI scoring rivals clinician agreement
PSO-Net estimates psoriasis severity from patient photos with ICC up to 87.8%, near the 88.1% human rater agreement.
Mamba model out-scores 8 methods on hyperspectral super-resolution
Local 3D windows and spectral similarity ordering lift PSNR on Chikusei and Houston.
RieszNet rivals 3D U-Net on cracks with 7,153 parameters
A scale-equivariant architecture built on Riesz transforms matches fine-tuned U-Net recall on multiscale cracks with far fewer parameters.
26.4M-parameter SAM student matches the original on most medical sets
The 26.4M student keeps Dice close to the 632M original, a step toward on-device use.
· “Efficient Knowledge Distillation of SAM for Medical Image Segmentation”
Brain activity alone reveals which movie clip a person watched
Surface vision transformers generalize to new viewers and unseen clips, with no per-person calibration.
Video transformer hits 98.6% accuracy on Alzheimer's MRI
Treating 3D brain scans as video helps an AI separate Alzheimer's, MCI, and normal aging.
· “Leveraging Video Vision Transformer for Alzheimer's Disease Diagnosis from 3D Brain MRI”
Attention module lifts small-lesion Dice across U-Net variants
Adding cross-scale attention to skip connections improves small stroke lesion segmentation in ATLAS v2.0.
· “Stroke Lesion Segmentation using Multi-Stage Cross-Scale Attention”
More recording hours, not more subjects, drive decoding gains
Across 84 subjects and 498 hours, decoding rises log-linearly with no plateau; per-person hours beat adding people.
Deep learning measures foam pores in seconds
Matches manual pore-size and anisotropy measurements for four foam families in about seven seconds per image.
· “On the use of neural networks for the structural characterization of polymeric porous materials”
Sparse objects tolerate more lensless image mixing
A decoder-free Fisher-information analysis predicts when aggressive multiplexing is safe under shot noise.
Diffusion restoration works without knowing the degradation model
A learned invertible network supplies data consistency, so no analytic degradation model is needed for blind restoration.
Point cloud quality scored with zero point cloud labels
Image-trained network scores unlabeled point clouds, beating older transfer methods by up to 40 percent.
Straighter optic nerves track thinner retinal layers
A new MRI orbital marker, ILPP distance, links globe position and size to axonal health across 18,000 eyes.
Diffusion model extends chest CT scans to show liver and kidneys
By filling in missing axial slices from overlapping CT datasets, old lung-screening scans become multi-organ volumes.
· “Beyond the Lungs: Extending the Field of View in Chest CT with Latent Diffusion Models”
Enhancement network hits 24.75 dB PSNR on X-rays
Per-pixel brightness and contrast maps show clinicians why each region is adjusted.
· “Quality Enhancement of Radiographic X-ray Images by Interpretable Mapping”
RL agent cuts bit rate 25% for car detection
Per-block QP control in standard H.264 preserves detection quality at lower bit rates, and the policy transfers to other detectors.
· “RL-RC-DoT: A Block-level RL agent for Task-Aware Video Compression”
No-reference proxies rank restorers like true MSE and LPIPS
A learned degradation estimator powers the proxies and a consistency-guided diffusion restorer called ELAD.
· “Proxies for Distortion and Consistency with Applications for Real-World Image Restoration”
CNN-RNN reads GPR track scans faster than Faster R-CNN
Review argues synthetic radar data plus a lightweight neural network can spot fouled ballast and subgrade defects without image conversion.
· “Advanced technology in railway track monitoring using the GPR Technique: A Review”
One shared 3D feature space turns any sensor view into any other
CrossModalityDiffusion generates novel views across EO, LiDAR, and SAR from a few images, with no geometry ground truth.
· “CrossModalityDiffusion: Multi-Modal Novel View Synthesis with Unified Intermediate Representation”
MobileNet and BiGRU sweep near-perfect COVID-19 benchmarks
One review benchmarks 11 standard models on X-rays, coughs, and tweets; MobileNet wins image and audio, BiGRU wins text.
VGG19 pipeline blends multiple artistic styles into one image
Weighted content and style losses run the styles one after another, letting users tune each painting's influence.
· “Dynamic Neural Style Transfer for Artistic Image Generation using VGG19”
Temporal-aware diffusion sharpens accelerated heart and lung MRI
Joining image-domain and frequency-domain temporal priors keeps motion aligned at up to 10x acceleration and 17 radial spokes.
Self-pretrained mask autoencoder boosts 3D infarct Dice by 3.7%
Pretraining on 45,374 MRI studies with variable contrasts sharpens ViT-based infarct segmentation.
· “Self Pre-training with Adaptive Mask Autoencoders for Variable-Contrast 3D Medical Imaging”
Retinal center-surround filtering explains the expanding hole illusion
The same classical Difference of Gaussians filter makes a white hole shrink, tying a static image to early vision.
Tutorial shows how VAEs turn brain scans into interpretable variables
It derives the ELBO and reparameterization trick, showing latent variables reveal disease or collapse to average brain.
New loss beats cross-entropy while targeting the Bayes error
Sampling-based BOLT loss minimizes an upper bound on the minimum achievable error, matching or beating cross-entropy in tests.
· “Universal Training of Neural Networks to Achieve Bayes Optimal Classification Accuracy”
2D Gaussians outclass INR models for arbitrary-scale super-resolution
Feed-forward rendering of continuous Gaussians beats per-pixel MLP queries on quality and speed, ×2 to ×30.
· “Generalized and Efficient 2D Gaussian Splatting for Arbitrary-scale Super-Resolution”
Vision Transformer plus CBAM hits 93.46% on 39 skin lesion types
Merging five public skin-image sets into 39 balanced classes puts attention-guided transformers ahead of all baselines.
· “An Attention-Guided Deep Learning Approach for Classifying 39 Skin Lesion Types”
Two X-ray views yield labeled 3D bone models in one pass
A Swin-transformer network reconstructs 54 bone classes across four anatomies and reports top scores on nine DRR datasets.
· “Swin-X2S: Reconstructing 3D Shape from 2D Biplanar X-ray with Swin Transformers”
Survey maps five GAN families for underwater image restoration
If the survey is right, its taxonomy, datasets, and loss formulas give newcomers a reliable entry point.
· “Underwater Image Enhancement using Generative Adversarial Networks: A Survey”
Network splits one stained image into two organelle channels
One fluorescence channel plus AEMS-Net replaces two-channel sequential acquisition, halving staining and light exposure for live cells.
Enhancement trained for recognition nearly doubles segmentation mIoU
A 577k-parameter filter lifts NightCity mIoU from 18.4 to 34.4 and pose AP from 32.4 to 34.1, no retraining.
· “Recognition-Oriented Low-Light Image Enhancement based on Global and Pixelwise Optimization”
DenseNet achieves 97.64% accuracy in COVID chest X-ray multiclass test
Eight pre-trained networks were compared on 3,856 X-rays; binary precision hit 99.98%.
· “Comparison of Neural Models for X-ray Image Classification in COVID-19 Detection”
Gaussian blobs give 4D lung CT from a 1-minute scan
Joint blob-and-motion optimization matches prior-image methods and sets the lowest target shift error.
· “Spatiotemporal Gaussian Optimization for 4D Cone Beam CT Reconstruction from Sparse Projections”
SCConv nudges YOLOv9 brain-tumor detection up 0.3–0.5 points
A plug-in redundancy-reduction module edges out plain YOLOv9 and SE attention on two MRI datasets.
· “SCC-YOLO: An Improved Object Detector for Assisting in Brain Tumor Diagnosis”
Logit clustering boosts LLM ultrasound-quality accuracy by 10.11%
New expert-labeled benchmark of 11,572 ultrasound images shows MLLMs need this post-processing to counter quality bias.
· “Ultrasound-QBench: Can LLMs Aid in Quality Assessment of Ultrasound Imaging?”
New MRI model hits 87.49% Dice for stroke lesions
Combining DWI, ADC, and enhanced DWI beats the top ISLES 2022 scores by 5.4 points.
· “Deep Learning-Driven Segmentation of Ischemic Stroke Lesions Using Multi-Channel MRI”
A simple network reveals where a gamma probe is sensing
Surgeons could see the probe-tissue intersection live, replacing guesswork from audible gamma counts.
