Set Of Natural Images Research Articles

Research on the Registration of Aerial Images of Cyclobalanopsis Natural Forest Based on Optimized Fast Sample Consensus Point Matching with SIFT Features

The effective management and conservation of forest resources hinge on accurate monitoring. Nonetheless, individual remote-sensing images captured by low-altitude unmanned aerial vehicles (UAVs) fail to encapsulate the entirety of a forest’s characteristics. The application of image-stitching technology to high-resolution drone imagery facilitates a prompt evaluation of forest resources, encompassing quantity, quality, and spatial distribution. This study introduces an improved SIFT algorithm designed to tackle the challenges of low matching rates and prolonged registration times encountered with forest images characterized by dense textures. By implementing the SIFT-OCT (SIFT omitting the initial scale space) approach, the algorithm bypasses the initial scale space, thereby reducing the number of ineffective feature points and augmenting processing efficiency. To bolster the SIFT algorithm’s resilience against rotation and illumination variations, and to furnish supplementary information for registration even when fewer valid feature points are available, a gradient location and orientation histogram (GLOH) descriptor is integrated. For feature matching, the more computationally efficient Manhattan distance is utilized to filter feature points, which further optimizes efficiency. The fast sample consensus (FSC) algorithm is then applied to remove mismatched point pairs, thus refining registration accuracy. This research also investigates the influence of vegetation coverage and image overlap rates on the algorithm’s efficacy, using five sets of Cyclobalanopsis natural forest images. Experimental outcomes reveal that the proposed method significantly reduces registration time by an average of 3.66 times compared to that of SIFT, 1.71 times compared to that of SIFT-OCT, 5.67 times compared to that of PSO-SIFT, and 3.42 times compared to that of KAZE, demonstrating its superior performance.

Foundation models in gastrointestinal endoscopic AI: Impact of architecture, pre-training approach and data efficiency

Pre-training deep learning models with large data sets of natural images, such as ImageNet, has become the standard for endoscopic image analysis. This approach is generally superior to training from scratch, due to the scarcity of high-quality medical imagery and labels. However, it is still unknown whether the learned features on natural imagery provide an optimal starting point for the downstream medical endoscopic imaging tasks. Intuitively, pre-training with imagery closer to the target domain could lead to better-suited feature representations. This study evaluates whether leveraging in-domain pre-training in gastrointestinal endoscopic image analysis has potential benefits compared to pre-training on natural images.To this end, we present a dataset comprising of 5,014,174 gastrointestinal endoscopic images from eight different medical centers (GastroNet-5M), and exploit self-supervised learning with SimCLRv2, MoCov2 and DINO to learn relevant features for in-domain downstream tasks. The learned features are compared to features learned on natural images derived with multiple methods, and variable amounts of data and/or labels (e.g. Billion-scale semi-weakly supervised learning and supervised learning on ImageNet-21k). The effects of the evaluation is performed on five downstream data sets, particularly designed for a variety of gastrointestinal tasks, for example, GIANA for angiodyplsia detection and Kvasir-SEG for polyp segmentation.The findings indicate that self-supervised domain-specific pre-training, specifically using the DINO framework, results into better performing models compared to any supervised pre-training on natural images. On the ResNet50 and Vision-Transformer-small architectures, utilizing self-supervised in-domain pre-training with DINO leads to an average performance boost of 1.63% and 4.62%, respectively, on the downstream datasets. This improvement is measured against the best performance achieved through pre-training on natural images within any of the evaluated frameworks.Moreover, the in-domain pre-trained models also exhibit increased robustness against distortion perturbations (noise, contrast, blur, etc.), where the in-domain pre-trained ResNet50 and Vision-Transformer-small with DINO achieved on average 1.28% and 3.55% higher on the performance metrics, compared to the best performance found for pre-trained models on natural images.Overall, this study highlights the importance of in-domain pre-training for improving the generic nature, scalability and performance of deep learning for medical image analysis. The GastroNet-5M pre-trained weights are made publicly available in our repository: huggingface.co/tgwboers/GastroNet-5M_Pretrained_Weights.

The universal law of generalization holds for naturalistic stimuli.

Shepard's universal law of generalization is a remarkable hypothesis about how intelligent organisms should perceive similarity. In its broadest form, the universal law states that the level of perceived similarity between a pair of stimuli should decay as a concave function of their distance when embedded in an appropriate psychological space. While extensively studied, evidence in support of the universal law has relied on low-dimensional stimuli and small stimulus sets that are very different from their real-world counterparts. This is largely because pairwise comparisons-as required for similarity judgments-scale quadratically in the number of stimuli. We provide strong evidence for the universal law in a naturalistic high-dimensional regime by analyzing an existing data set of 214,200 human similarity judgments and a newly collected data set of 390,819 human generalization judgments (N = 2,406 U.S. participants) across three sets of natural images. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Trypophobia, skin disease, and the visual discomfort of natural textures

In the last decade, the behavioral sciences have described the phenomenon of trypophobia, which is the discomfort felt by some individuals when viewing images containing clusters of bumps or holes. One evolutionary hypothesis for this phenomenon is that this visual discomfort represents an adaptation which helps organisms avoid skin disease and/or ectoparasites. Even though trypophobic imagery and disease imagery are both examples of visual textures, to date there has been no comparison of the visual discomfort elicited by these two specific kinds of textures within the larger context of the visual comfort elicited by natural texture imagery more generally. In the present study, we administered the Trypophobia Questionnaire (TQ) and recorded the visual comfort ratings elicited by a large set of standard natural texture images, including several trypophobic and skin disease images. In two independent samples, we found that while all observers find skin diseases uncomfortable to view, only those scoring high on the TQ rated trypophobic imagery equally uncomfortable. Comparable effects were observed using both standard ANOVA procedures as well as linear mixed effects modeling. Comparing the ratings of both high-TQ and low-TQ participants to the standard textures, we find remarkably consistent rank-order preferences, with the most unpleasant textures (as rated by both groups) exhibiting qualitative similarities to trypophobic imagery. However, we also find that low-level image statistics which have been previously shown to affect visual comfort are poor predictors of the visual comfort elicited by natural textures, including trypophobic and disease imagery. Our results suggest that a full understanding of the visual comfort elicited by natural textures, including those arising from skin disease, will ultimately depend upon a better understanding of cortical areas specialized for the perception of surface and material properties, and how these visual regions interact with emotional brain areas to evoke appropriate behavioral responses, like disgust.

SFA-Net: A Selective Features Absorption Network for Object Detection in Rainy Weather Conditions.

In recent years, object detection approaches using deep convolutional neural networks (CNNs) have derived major advances in normal images. However, such success is hardly achieved with rainy images due to lack of visibility. Aiming to bridge this gap, in this article, we present a novel selective features absorption network (SFA-Net) to improve the performance of object detection not only in rainy weather conditions but also in favorable weather conditions. SFA-Net accomplishes this objective by utilizing three subnetworks, where the feature selection subnetwork is concatenated with the object detection subnetwork through the feature absorption subnetwork to form a unified model. To promote further advancement in object detection impaired by rain, we propose a large-scale rainy image dataset, named srRain, which contains both synthetic rainy images and real-world rainy images for training and testing purposes. srRain is comprised of 25 900 rainy images depicting diverse driving scenarios in the presence of rain with a total of 181 164 instances interpreting five common object categories. Experimental results display that our SFA-Net reaches the highest mean average precision (mAP) of 77.53% on a normal image set, 62.52% on a synthetic rainy image set, 37.34% on a collected natural rainy image set, and 32.86% on a published real rainy image set, surpassing current state-of-the-art object detectors and the combination of image deraining and object detection models while retaining a high speed.

Establishing functional homology across species using a common set of natural images

Establishing functional homology across species using a common set of natural images

An active foveated gaze prediction algorithm based on a Bayesian ideal observer

Predicting human eye movements is a crucial task for understanding human behavior and has numerous applications in machine vision. Most current models for predicting eye movements are data-driven and require large datasets of recorded eye movements, which can be expensive and time-consuming to collect. In this paper, we present a novel theory-based model for predicting eye movements in a foveated visual system that maximizes information gain at each fixation. Our model uses a region-proposal network and eccentricity-based max pooling to account for the loss of detail in peripheral vision. We apply our model to predict human fixations in a visual search task for objects in real-world scenes. Unlike data-driven models, our model does not require training on large eye movement datasets and can generalize to any set of natural images and targets. We evaluate the generalization capability of our model by demonstrating its results on two publicly available visual search datasets, Ehinger and COCO-search18, without any further training on those datasets. Our model outperforms or performs comparably to data-driven models that are directly trained on human eye movement datasets.

Objective Prediction of Human Visual Acuity Using Image Quality Metrics

This work addresses the objective prediction of human uncorrected decimal visual acuity, an unsolved challenge due to the contribution of both physical and neural factors. An alternative approach to assess the image quality of the human visual system can be addressed from the image and video processing perspective. Human tolerance to image degradation is quantified by mean opinion scores, and several image quality assessment algorithms are used to maintain, control, and improve the quality of processed images. The aberration map of the eye is used to obtain the degraded theoretical image from a set of natural images. The amount of distortion added by the eye to the natural image was quantified using different image processing metrics, and the correlation between the result of each metric and subjective visual acuity was assessed. The correlation obtained for a model based on a linear combination of the normalized mean square error metric and the feature similarity index metric was very good. It was concluded that the proposed method could be an objective way to determine subjects’ monocular and uncorrected decimal visual acuity with low uncertainty.

Low-level tuning biases in higher visual cortex reflect the semantic informativeness of visual features.

Representations of visual and semantic information can overlap in human visual cortex, with the same neural populations exhibiting sensitivity to low-level features (orientation, spatial frequency, retinotopic position) and high-level semantic categories (faces, scenes). It has been hypothesized that this relationship between low-level visual and high-level category neural selectivity reflects natural scene statistics, such that neurons in a given category-selective region are tuned for low-level features or spatial positions that are diagnostic of the region's preferred category. To address the generality of this "natural scene statistics" hypothesis, as well as how well it can account for responses to complex naturalistic images across visual cortex, we performed two complementary analyses. First, across a large set of rich natural scene images, we demonstrated reliable associations between low-level (Gabor) features and high-level semantic categories (faces, buildings, animate/inanimate objects, small/large objects, indoor/outdoor scenes), with these relationships varying spatially across the visual field. Second, we used a large-scale functional MRI dataset (the Natural Scenes Dataset) and a voxelwise forward encoding model to estimate the feature and spatial selectivity of neural populations throughout visual cortex. We found that voxels in category-selective visual regions exhibit systematic biases in their feature and spatial selectivity, which are consistent with their hypothesized roles in category processing. We further showed that these low-level tuning biases are not driven by selectivity for categories themselves. Together, our results are consistent with a framework in which low-level feature selectivity contributes to the computation of high-level semantic category information in the brain.

Seismic full-waveform inversion based on superresolution for high-precision prediction of reservoir parameters

The frequency band limitation of seismic data limits the resolution of full-waveform inversion (FWI) results. In addition, the high computational cost seriously affects the practical application of FWI. To alleviate these concerns, an FWI method based on superresolution (SR-FWI) is developed to improve the prediction efficiency and accuracy of the reservoir parameters. A channel attention mechanism is introduced for the multifrequency characteristics of the model images. A constrained residual channel attention network (CRCAN) is built for superresolution (SR) by adding structural constraints to the loss function of a deep learning network. A total of 65,000 sets of geologic models and natural images constitute the network training data, 90% of which are used for training with the rest used for testing. The iterative calculation for FWI is time-consuming; hence, SR is applied to the iterative process to reduce the number of iterations and accelerate the model update. Low-resolution images along with the synthetic and field data are used for the evaluation of the CRCAN and SR-FWI algorithms, respectively. The test results find that CRCAN can effectively improve the image resolution, whereas SR-FWI is beneficial due to its high efficiency and precision, especially in predicting the stratum edge and small-scale anomalies. Therefore, SR-FWI is a powerful means of reservoir static and dynamic detection and can provide high-resolution information for projects, such as resource development and CO2 storage.

Inter-individual deep image reconstruction via hierarchical neural code conversion

The sensory cortex is characterized by general organizational principles such as topography and hierarchy. However, measured brain activity given identical input exhibits substantially different patterns across individuals. Although anatomical and functional alignment methods have been proposed in functional magnetic resonance imaging (fMRI) studies, it remains unclear whether and how hierarchical and fine-grained representations can be converted between individuals while preserving the encoded perceptual content. In this study, we trained a method of functional alignment called neural code converter that predicts a target subject’s brain activity pattern from a source subject given the same stimulus, and analyzed the converted patterns by decoding hierarchical visual features and reconstructing perceived images. The converters were trained on fMRI responses to identical sets of natural images presented to pairs of individuals, using the voxels on the visual cortex that covers from V1 through the ventral object areas without explicit labels of the visual areas. We decoded the converted brain activity patterns into the hierarchical visual features of a deep neural network using decoders pre-trained on the target subject and then reconstructed images via the decoded features. Without explicit information about the visual cortical hierarchy, the converters automatically learned the correspondence between visual areas of the same levels. Deep neural network feature decoding at each layer showed higher decoding accuracies from corresponding levels of visual areas, indicating that hierarchical representations were preserved after conversion. The visual images were reconstructed with recognizable silhouettes of objects even with relatively small numbers of data for converter training. The decoders trained on pooled data from multiple individuals through conversions led to a slight improvement over those trained on a single individual. These results demonstrate that the hierarchical and fine-grained representation can be converted by functional alignment, while preserving sufficient visual information to enable inter-individual visual image reconstruction.

A multi-model data-fusion based deep transfer learning for improved remaining useful life estimation for IIOT based systems

Remaining useful life (RUL) estimation, a key component in predictive maintenance (PdM), aims to reduce maintenance cycles in the prognostic health of mechanical equipment(s). Research directions using deep-learning-based RUL estimation often suffer from limited availability of degradation signals resulting inaccurate predictions. Until now, state-of-the-art models trained on large sets of natural images to classify objects have not been re-used to improve regression-based RUL estimation accuracies of mechanical equipment. Actually, this is a rarely researched topic in PdM. Inspired by transfer learning, we showcase that the knowledge learned by popular pre-trained models can be transferred to improve industrial machinery-based-maintenance decision-making. Accordingly, this paper proposes a novel multi-model data-fusion-based deep transfer learning (MMF-DTL) framework for improved RUL estimation of rolling bearings through degradation images (DI) and pre-trained deep convolutional neural networks (CNNs). After procuring the degradation signals, we obtain DIs incorporating sufficient deterioration information using Markov Transition Fields. Next, these DIs are input into a DTL network comprising three pre-trained CNNs, i.e., DenseNet201, VGG16, and ResNet50, designed in a parallel fashion, where each constituent fine-tunes a different count of layers. Subsequently, features extracted from each component model are fused in a weighted manner and passed onto several fully connected layers. The effectiveness of the proposed framework is validated using the PHM Challenge 2012 bearing degradation dataset. Compared to several state-of-the-art works, our approach improves ∼12.57% on error rate and ∼26.04% on MAE, suggesting it is practically feasible to grasp transferable attributes from a general-purpose related dataset to another with minimal dataset size.

Automated Determination of Left Ventricular Function Using Electrocardiogram Data in Patients on Maintenance Hemodialysis.

Left ventricular ejection fraction is disrupted in patients on maintenance hemodialysis and can be estimated using deep learning models on electrocardiograms. Smaller sample sizes within this population may be mitigated using transfer learning. We identified patients on hemodialysis with transthoracic echocardiograms within 7 days of electrocardiogram using diagnostic/procedure codes. We developed four models: (1) trained from scratch in patients on hemodialysis, (2) pretrained on a publicly available set of natural images (ImageNet), (3) pretrained on all patients not on hemodialysis, and (4) pretrained on patients not on hemodialysis and fine-tuned on patients on hemodialysis. We assessed the ability of the models to classify left ventricular ejection fraction into clinically relevant categories of ≤40%, 41% to ≤50%, and >50%. We compared performance by area under the receiver operating characteristic curve. We extracted 705,075 electrocardiogram:echocardiogram pairs for 158,840 patients not on hemodialysis used for development of models 3 and 4 and n=18,626 electrocardiogram:echocardiogram pairs for 2168 patients on hemodialysis for models 1, 2, and 4. The transfer learning model achieved area under the receiver operating characteristic curves of 0.86, 0.63, and 0.83 in predicting left ventricular ejection fraction categories of ≤40% (n=461), 41%-50% (n=398), and >50% (n=1309), respectively. For the same tasks, model 1 achieved area under the receiver operating characteristic curves of 0.74, 0.55, and 0.71, respectively; model 2 achieved area under the receiver operating characteristic curves of 0.71, 0.55, and 0.69, respectively, and model 3 achieved area under the receiver operating characteristic curves of 0.80, 0.51, and 0.77, respectively. We found that predictions of left ventricular ejection fraction by the transfer learning model were associated with mortality in a Cox regression with an adjusted hazard ratio of 1.29 (95% confidence interval, 1.04 to 1.59). A deep learning model can determine left ventricular ejection fraction for patients on hemodialysis following pretraining on electrocardiograms of patients not on hemodialysis. Predictions of low ejection fraction from this model were associated with mortality over a 5-year follow-up period. This article contains a podcast at https://www.asn-online.org/media/podcast/CJASN/2022_06_06_CJN16481221.mp3.

Could simplified stimuli change how the brain performs visual search tasks? A deep neural network study.

Visual search is a complex behavior influenced by many factors. To control for these factors, many studies use highly simplified stimuli. However, the statistics of these stimuli are very different from the statistics of the natural images that the human visual system is optimized by evolution and experience to perceive. Could this difference change search behavior? If so, simplified stimuli may contribute to effects typically attributed to cognitive processes, such as selective attention. Here we use deep neural networks to test how optimizing models for the statistics of one distribution of images constrains performance on a task using images from a different distribution. We train four deep neural network architectures on one of three source datasets—natural images, faces, and x-ray images—and then adapt them to a visual search task using simplified stimuli. This adaptation produces models that exhibit performance limitations similar to humans, whereas models trained on the search task alone exhibit no such limitations. However, we also find that deep neural networks trained to classify natural images exhibit similar limitations when adapted to a search task that uses a different set of natural images. Therefore, the distribution of data alone cannot explain this effect. We discuss how future work might integrate an optimization-based approach into existing models of visual search behavior.

Deformable models for image segmentation: A critical review of achievements and future challenges

Image segmentation is a fundamental and tedious task of computer vision. Because of inherent noise and intensity inhomogeneity in real-world images, it remains a difficult problem in practical applications such as image analysis, scene understanding, object detection, and many others. Several mathematical models proposed for image segmentation in the past few decades with an effective policy. Among these models, deformable models earned more attention and are widely used techniques due to their accuracy, efficiency, and robust effectiveness. This study reviews and compares various deformable models for the segmentation of digital images available in the literature. First, for a comprehensive study, these deformable models are classified into two classes such as direct partial differential equation (PDE) based approaches and variational based approaches. Beside this, variational based approaches are further classified into parametric and geometric models. Their advantages as well as shortcomings are discussed in detail from an objective viewpoint. Then, to check the robustness of various discussed classes of deformable models, a set of synthetic, natural, and real medical images are considered along with inhomogeneity and noise. Also, to measure the segmentation accuracy, different quantitative metrics based on contour and region are utilized. Numerical experiments with different classes of images, reveal that both direct PDE based and variation based models perform well in clean and noisy images. Whereas, the variational based approaches are superior for images having intensity inhomogeneity with high computational time. The qualitative and quantitative investigations confirm that the subtle change in model assumptions can have a significant impact on segmentation.

341 An intracranial EEG map of naturalistic images in the human brain

OBJECTIVES/GOALS: Our overall goal is to identify the processes used by the human visual system to encode visual stimuli into perceptual representations. In this project, our objective is (i) to collect a dataset of human neural activity in response to 1000 naturalistic color images and (ii) to determine how image parameters drive different parts of the human brain. METHODS/STUDY POPULATION: We recorded iEEG data in 4 human subjects who had been implanted for epilepsy monitoring. Each subject was presented 10 sets of 100 naturalistic stimuli, taken from the Natural Scenes Dataset (Allen et al., 2021), on a screen for 1 second each with 1 second rest intervals between stimuli. The subjects were instructed to fixate on a red dot at the center of the screen and were prompted to recall whether they had seen 3 additional test stimuli at the end of each set to encourage attentiveness. We calculated significant neural responses at each electrode by comparing evoked potentials and high frequency power changes during each stimulus vs. rest. Electrodes with significant responses were then mapped to anatomic locations in each subjects brain and then collectively to a standard brain. RESULTS/ANTICIPATED RESULTS: The natural image set elicited significant evoked potentials and high frequency responses at electrodes in each subject. Response latencies, from 80 to 300 ms after stimulus onset, portrayed the evolution of visual processing along the visual pathways, through key sites such as the early visual cortex, ventral temporal cortex, intraparietal sulcus, and frontal eye field. These responses differed significantly from those elicited by simple patterns, which drove early visual cortex but less so in later regions. DISCUSSION/SIGNIFICANCE: These data show that the human brain responds differently to more complex images. Determining the human brains response to naturalistic images is essential for encoding models that describe the processing in the human visual system. These models may further future efforts for electrical neurostimulation therapies such as for restoring vision.

Introducing the Prototypical Stimulus Characteristics Toolbox: Protosc

Many studies use different categories of images to define their conditions. Since any difference between these categories is a valid candidate to explain category-related behavioral differences, knowledge about the objective image differences between categories is crucial for the interpretation of the behaviors. However, natural images vary in many image features and not every feature is equally important in describing the differences between the categories. Here, we provide a methodological approach to find as many of the image features as possible, using machine learning performance as a tool, that have predictive value over the category the images belong to. In other words, we describe a means to find the features of a group of images by which the categories can be objectively and quantitatively defined. Note that we are not aiming to provide a means for the best possible decoding performance; instead, our aim is to uncover prototypical characteristics of the categories. To facilitate the use of this method, we offer an open-source, MATLAB-based toolbox that performs such an analysis and aids the user in visualizing the features of relevance. We first applied the toolbox to a mock data set with a ground truth to show the sensitivity of the approach. Next, we applied the toolbox to a set of natural images as a more practical example.

SNR Analysis for Quantitative Comparison of Line Detection Methods

The need for line detection in images is growing rapidly due to its importance in many image processing applications. The selection of an appropriate line detection method is essential for accurate detection of line pixels, but few studies provide an analytical basis for selecting a specific line detection method. In this study, to solve the problem, a method to analytically determine the signal-to-noise ratio (SNR) of line detection methods is proposed. Three line detection methods were selected for comparison: edge-detection (ED)-based, second derivative (SD)-based, and the sum of gradient angle differences (SGAD)-based line detection methods. Then, this study quantifies the SNR of the three line detectors through error propagation and signal noise coupling. In addition, the derived SNRs are graphically visualized to explicitly compare the performance of line detectors. Then, the quantified SNRs were validated by showing that they are highly correlated with the completeness and correctness observed in the experiment with a set of natural images. The experimental results show that the proposed SNR analysis can be used to select or design a suitable line detector.

Critical information thresholds underlying generic and familiar face categorisation at the same face encounter

Seeing a face in the real world provokes a host of automatic categorisations related to sex, emotion, identity, and more. Such individual facets of human face recognition have been extensively examined using overt categorisation judgements, yet their relative informational dependencies during the same face encounter are comparatively unknown. Here we used EEG to assess how increasing access to sensory input governs two ecologically relevant brain functions elicited by seeing a face: Distinguishing faces and nonfaces, and recognising people we know. Observers viewed a large set of natural images that progressively increased in either image duration (experiment 1) or spatial frequency content (experiment 2). We show that in the absence of an explicit categorisation task, the human brain requires less sensory input to categorise a stimulus as a face than it does to recognise whether that face is familiar. Moreover, where sensory thresholds for distinguishing faces/nonfaces were remarkably consistent across observers, there was high inter-individual variability in the lower informational bound for familiar face recognition, underscoring the neurofunctional distinction between these categorisation functions. By i) indexing a form of face recognition that goes beyond simple low-level differences between categories, and ii) tapping multiple recognition functions elicited by the same face encounters, the information minima we report bear high relevance to real-world face encounters, where the same stimulus is categorised along multiple dimensions at once. Thus, our finding of lower informational requirements for generic vs. familiar face recognition constitutes some of the strongest evidence to date for the intuitive notion that sensory input demands should be lower for recognising face category than face identity.

Visual memorability in the absence of semantic content

What determines how well people remember images? Most past research has explored properties of the people doing the remembering — such as their age, emotional state, or individual capacity. However, recent work has also characterized memorability — the likelihood of an image being remembered across observers. But what makes some images more memorable than others? Part of the answer must surely involve the meanings of the images, but here we ask whether this is the entire story: is there also purely visual memorability, driven not by semantic content but by perceptual features per se? We isolated visual memorability in an especially direct manner — by eliminating semantic content while retaining many visual properties. We did so by transforming a set of natural scene images using phase scrambling, and then testing memorability for both intact and scrambled images in independent samples. Across several experiments, observers saw sequences of images and responded anytime they saw a repeated image. We found reliable purely visual memorability at the temporal scales of both short-term memory (2–15 s) and longer-term memory (several minutes), and this could not be explained by the extent to which people could generate semantic labels for some scrambled images. Collectively, these results suggest that the memorability of images is a function not only of what they mean, but also of how they look in the first place.