Abstract Plasma membrane homeostasis requires balanced exocytosis and endocytosis, yet their coordination at the single-event level in non-neuronal cells is unresolved. We present Shape2Fate, a morphology-aware deep-learning pipeline that detects, tracks, and classifies individual exocytic and endocytic carriers in live-cell total internal reflection fluorescence structured illumination microscopy (TIRF-SIM) movies at ∼100 nm resolution. Trained on synthetic data and exploiting carrier shape evolution rather than fluorescence intensity, Shape2Fate achieves expert-level tracking and outcome classification across diverse cell types, imaging conditions, and microscope platforms. Applying Shape2Fate to constitutive secretion and insulin-stimulated GLUT4 exocytosis in adipocytes, we uncover two opposing exo-endocytic coupling architectures: exocytic fusion locally nucleates de novo clathrin-coated pits, whereas GLUT4 vesicles target pre-existing pits for rapid cargo capture. These findings establish that the spatial rules governing exo-endocytic coordination are not universal but are pathway-specific. Shape2Fate is openly available, enabling direct event-level mechanistic dissection of exo-endocytic coordination across pathways in living cells.

The paper deals with the recognition of blurred color/multispectral images directly without any deblurring. We present a general theory of invariants of multispectral images with respect to blur. The paper is a significant nontrivial extension of the recent theory of blur invariants of graylevel images. The main original contribution of the paper lies in introducing cross-channel blur invariants in Fourier domain. We also developed an algorithm for their stable and fast calculation in the moment domain. Moreover, the cross-channel invariants can be found for blurs for which single-channel invariants do not exist. The experiments on simulated and real data demonstrate that incorporating the new cross-channel invariants significantly improves the recognition power and surpasses other existing approaches. The outlook for a possible implementation of the blur invariants into neural networks is briefly sketched in the conclusion.

The paper deals with the recognition of color images that were degraded by blur and linear channel mixing with an unknown matrix. We propose invariants to this compound degradation in the Fourier domain by means of projection operators. We show an equivalent definition of the invariants in the image domain by means of moment expansion, which is more stable and faster when evaluated numerically. The experiment demonstrates a good performance of the invariants and their advantages over a recently proposed method of a similar nature. Possible modifications of the proposed invariants are discussed in the conclusion.

Segmenting a new biomedical dataset usually means a domain-specific model trained on substantial annotation, or a foundation model steered at inference time. We present Exemplar, a few-shot segmenter that fuses a frozen DINOv3 backbone with a fixed bank of classical native-resolution filter responses in one lightweight head, fitted from the support masks alone. In the few-mask, native-resolution regime, classical priors and frozen self-supervised features are complementary: fused in one head, a single fixed configuration spans eleven biomedical imaging datasets. Under the same head, the classical bank alone reaches 0.693 on the eleven-dataset panel, scored by foreground intersection-over-union or centreline Dice, and the frozen features alone 0.672; the bank leads on seven of the eleven and the features on the rest, and fused they reach 0.782. Against five forward-pass few-shot methods, Exemplar leads in 54 of 55 method-dataset comparisons, 52 of them significant after Holm correction. From a single annotated mask it reaches 0.703 on the same panel, against 0.682 for a from-scratch nnU-Net trained on that same mask. At eight masks nnU-Net overtakes it on the panel mean, chiefly on centreline agreement, but takes 16-77x longer to fit.

This work addresses inverse problems in image restoration, focusing on recovering high-quality images from degraded observations, a critical task in fields like microscopy and digital photography. We examine both traditional variational methods and modern deep learning techniques, highlighting hybrid approaches that merge mathematical modeling with data-driven learning. Classical model-based methods use explicit regularization, like total variation, to incorporate prior knowledge and stabilize the inversion process. Meanwhile, deep learning approaches, both supervised and self-supervised, leverage implicit regularization, where network architectures capture and learn prior information from data. We present our recent advancements in this field and discuss the effectiveness of these complementary approaches in solving complex image restoration problems in theory and practice.

We propose a novel approach to enhance image demosaicking algorithms using implicit neural representations (INR). Our method employs a multi-layer perceptron to encode RGB images, combining original Bayer measurements with an initial estimate from existing demosaicking methods to achieve superior reconstructions. A key innovation is the integration of two loss functions: a Bayer loss for fidelity to sensor data and a complementary loss that regularizes reconstruction using interpolated data from the initial estimate. This combination, along with INR's inherent ability to capture fine details, enables high-fidelity reconstructions that incorporate information from both sources. Furthermore, we demonstrate that INR can effectively correct artifacts in state-of-the-art demosaicking methods when input data diverge from the training distribution, such as in cases of noise or blur. This adaptability highlights the transformative potential of INR-based demosaicking, offering a robust solution to this challenging problem.

Accurately measuring the duration of actors’ presence in videos is a challenging task that goes beyond actor recognition. We propose the STAR pipeline, the new model designed to analyze the time performers appear on screen across diverse video content, including movies and TV shows. The proposed model has been successfully deployed and tested by the Czech TV infrastructure provider. Our pipeline uses machine learning techniques for shot detection, face detection, tracking, recognition, and introduces a novel shot-based method for calculating screen time. We present extensive experiments proving the robustness and real-time performance of our approach. Alongside the pipeline, we introduce the STAR dataset to address the need for high-quality benchmarks in evaluating screen time models, now available for download.

CNNs exhibit inherent equivariance to image translation, leading to efficient parameter and data usage, faster learning, and improved robustness. The concept of translation equivariant networks has been successfully extended to rotation transformation using group convolution for discrete rotation groups and harmonic functions for the continuous rotation group encompassing 360^∘. We explore the compatibility of the SA mechanism with full rotation equivariance, in contrast to previous studies that focused on discrete rotation. We introduce the Harmformer, a harmonic transformer with a convolutional stem that achieves equivariance for both translation and continuous rotation. Accompanied by an end-to-end equivariance proof, the Harmformer not only outperforms previous equivariant transformers, but also demonstrates inherent stability under any continuous rotation, even without seeing rotated samples during training.

We introduce NeRD, a new demosaicking method for generating full-color images from Bayer patterns. Our approach leverages advancements in neural fields to perform demosaicking by representing an image as a coordinate-based neural network with sine activation functions. The inputs to the network are spatial coordinates and a low-resolution Bayer pattern, while the outputs are the corresponding RGB values. An encoder network, which is a blend of ResNet and U-net, enhances the implicit neural representation of the image to improve its quality and ensure spatial consistency through prior learning. Our experimental results demonstrate that NeRD outperforms traditional and state-of-the-art CNN-based methods and significantly closes the gap to transformer-based methods.

Blur is a common phenomenon in image acquisition that negatively influences the recognition rate of most classifiers. This paper studies the influence of image blurring of various types and sizes on the recognition rate achieved by a deep convolutional network. We confirm that the blur significantly decreases the performance if the network has been trained on clear images only. When the training set is augmented with blurred samples, the recognition rate becomes sufficiently high even if the blur in query images is of different size than the blur used for training. However, this is mostly not true if query images contain blur of a different type from the one used for training.

This paper proposes a novel approach to real-time automatic rim detection, classification, and inspection by combining traditional computer vision and deep learning techniques. At the end of every automotive assembly line, a quality control process is carried out to identify any potential defects in the produced cars. Common yet hazardous defects are related, for example, to incorrectly mounted rims. Routine inspections are mostly conducted by human workers that are negatively affected by factors such as fatigue or distraction. We have designed a new prototype to validate whether all four wheels on a single car match in size and type. Additionally, we present three comprehensive open-source databases, CWD1500, WHEEL22, and RB600, for wheel, rim, and bolt detection, as well as rim classification, which are free-to-use for scientific purposes.

Visual data, such as images and videos, are frequently used as evidence in court trials. If the data quality is insufficient to convince the court, a carefully tailored data processing algorithm supported with expert’s opinion is necessary. We present two real cases from our forensic expertise practice, in which we demonstrate a successful application of video superresolution that helped to convict offenders. The most important feature of image processing algorithms to be legally accepted by the court, is to rule out artifacts with realistic details, which are known to appear for example in deep learning methods.

Blur is an image degradation that is difficult to remove. Invariants with respect to blur offer an alternative way of a~description and recognition of blurred images without any deblurring. In this paper, we present an original unified theory of blur invariants. Unlike all previous attempts, the new theory does not require any prior knowledge of the blur type. The invariants are constructed in the Fourier domain by means of orthogonal projection operators and moment expansion is used for efficient and stable computation. It is shown that all blur invariants published earlier are just particular cases of this approach. Experimental comparison to concurrent approaches shows the advantages of the proposed theory.

The widespread popularity of equivariant networks underscores the significance of parameter efficient models and effective use of training data. At a time when robustness to unseen deformations is becoming increasingly important, we present H-NeXt, which bridges the gap between equivariance and invariance. H-NeXt is a parameter-efficient roto-translation invariant network that is trained without a single augmented image in the training set. Our network comprises three components: an equivariant backbone for learning roto-translation independent features, an invariant pooling layer for discarding roto-translation information, and a classification layer. H-NeXt outperforms the state of the art in classification on unaugmented training sets and augmented test sets of MNIST and CIFAR-10.

Classical blur models are based on simplifying assumptions, namely shift-equivariance and circular boundary condition (CBC), that rarely hold in practice. Shift-equivariance means that a shift of the input produces the same shift of the output, which implies that blur is spatially invariant and image aliasing is not present. The CBC assumes that the image is rectangular and periodically repeating. Discrepancies between simplified models and real blurred observations cause strong artifacts in image restoration. The common remedy is to increase the model complexity and remove simplifying assumptions. However, this also brings extra computational complexity to the restoration task. We present spectral pre-adaptation (SPA) that pre-processes blurred images so they can be restored using fast standard deconvolution algorithms suitable for simplified models. The SPA serves as a connector between classical deconvolution methods and a variety of real observations involving blur. Experiments on simulated and real images show that standard deconvolution of SPA-interpolated images not only greatly reduces artifacts compared to direct deconvolution, but performs on a par with more complex restoration methods.

Clathrin-mediated endocytosis (CME) is the main mechanism by which mammalian cells control their cell surface proteome. Proper operation of the pivotal CME cargo-adaptor AP2 requires membrane-localised FCHO. Here, live-cell eTIRF-SIM shows that FCHO marks sites of clathrin- coated pit (CCP) initiation, which mature into uniform sized CCPs comprising a central patch of AP2 and clathrin corralled by an FCHO/Eps15 ring. We dissect the network of interactions between the FCHO interdomain-linker and AP2, which concentrates, orients, tethers and partially destabilizes closed AP2 at the plasma membrane. AP2’s subsequent membrane deposition drives its opening, which triggers FCHO displacement through steric competition with PtdIns4,5P2, clathrin, cargo and CME accessory factors. FCHO can now relocate toward a CCP’s outer edge to engage and activate further AP2s to drive CCP growth/maturation.125 character summary FCHO primes AP2 for CCV incorporation, a process that triggers FCHO release to enable activation/recruitment of further AP2s### Competing Interest StatementThe authors have declared no competing interest.

Images are fused to produce a composite image by combining key characteristics of the source images in image fusion. It makes the fused image better for human vision and machine vision. A novel procedure of Infrared (IR) and Visible (Vis) image fusion is proposed in this manuscript. The main challenges of feature level image fusion are that it will introduce artifacts and noise in the fused image. To preserve the meaningful information without adding artifacts from the source input images, weight map computed from Arithmetic optimization algorithm (AOA) is used for the image fusion process. In this manuscript, feature level fusion is performed after refining the weight maps using a weighted least square optimization (WLS) technique. Through this, the derived salient object details are merged into the visual image without introducing distortion. To affirm the validity of the proposed methodology simulation results are carried for twenty-one image data sets. It is concluded from the qualitative and quantitative experimental analysis that the proposed method works well for most of the image data sets and shows better performance than certain traditional existing models.

Structured illumination microscopy is a widely popular super-resolution technique for live cell imaging capable of surpassing the diffraction limit. Its temporal resolution is limited by the need to capture multiple low-resolution images to reconstruct a single high-resolution image. When observing rapid biological processes, the local movement between frames leads to the formation of reconstruction artifacts, which subsequently impair the data interpretation. We propose to include this type of movement in the definition of the image formation forward problem. The motion can then be estimated from the original data using optical flow, and the optimization problem is solved using the alternating direction method of multipliers. Our approach is tested against other reconstruction techniques on both synthetic and real biological data.

Clathrin-mediated endocytosis (CME) is the main mechanism by which mammalian cells control their cell surface proteome. Proper operation of the pivotal CME cargo adaptor AP2 requires membrane-localized Fer/Cip4 homology domain-only proteins (FCHO). Here, live-cell enhanced total internal reflection fluorescence–structured illumination microscopy shows that FCHO marks sites of clathrin-coated pit (CCP) initiation, which mature into uniform-sized CCPs comprising a central patch of AP2 and clathrin corralled by an FCHO/Epidermal growth factor potential receptor substrate number 15 (Eps15) ring. We dissect the network of interactions between the FCHO interdomain linker and AP2, which concentrates, orients, tethers, and partially destabilizes closed AP2 at the plasma membrane. AP2’s subsequent membrane deposition drives its opening, which triggers FCHO displacement through steric competition with phosphatidylinositol 4,5-bisphosphate, clathrin, cargo, and CME accessory factors. FCHO can now relocate toward a CCP’s outer edge to engage and activate further AP2s to drive CCP growth/maturation.

The field of blind image deblurring was for a long time dominated by Maximum-A-Posteriori methods seeking the optimal pair of sharp image-blur of a suitable functional. Recently, learning-based methods, especially those based on deep convolutional neural networks, are proving effective and are receiving increasing attention by the research community. In 2020, Ren et al. proposed a deblurring method called SelfDeblur which combines the model-driven approach of traditional MAP methods and the generative power of neural nets. The method is capable of producing very high-quality results, yet it inherits some problems of MAP methods, especially possible convergence to a wrong local optimum. In this paper we propose several easy-to-implement modifications of SelfDeblur, namely suitable initialization, multiscale processing, and regularization, that improve the average performance of the original method and decrease the probability of failure.