Search

Delineation of retinal vessels in fundus images is essential for detecting a range of eye disorders. An automated technique for vessel segmentation can assist clinicians and enhance the efficiency of the diagnostic process. Traditional methods fail to extract multiscale information, discard unnecessary information, and delineate thin vessels. In this paper, a novel residual U-Net architecture that incorporates multi-scale feature learning and effective attention is proposed to delineate the retinal vessels precisely. Since drop block regularization performs better than drop out in preventing overfitting, drop block was used in this study. A multi-scale feature learning module was added instead of a skip connection to learn multi-scale features. A novel effective attention block was proposed and integrated with the decoder block to obtain precise spatial and channel information. Experimental findings indicated that the proposed model exhibited outstanding performance in retinal vessel delineation. The sensitivities achieved for DRIVE, STARE, and CHASE_DB datasets were 0.8293, 0.8151 and 0.8084, respectively.

Hypertensive retinopathy (HR) results from the microvascular retinal changes triggered by hypertension, which is the most common leading cause of preventable blindness worldwide. Therefore, it is necessary to develop an automated system for HR detection and evaluation using retinal images. We aimed to propose an automated approach to identify and categorize the various degrees of HR severity. A new network called the spatial convolution module (SCM) combines cross-channel and spatial information, and the convolution operations extract helpful features. The present model is evaluated using publicly accessible datasets ODIR, INSPIREVR, and VICAVR. We applied the augmentation to artificially increase the dataset of 1200 fundus images. The different HR severity levels of normal, mild, moderate, severe, and malignant are finally classified with the reduced time when compared to the existing models because in the proposed model, convolutional layers run only once on the input fundus images, which leads to a speedup and reduces the processing time in detecting the abnormalities in the vascular structure. According to the findings, the improved SVM had the highest detection and classification accuracy rate in the vessel classification with an accuracy of 98.99% and completed the task in 160.4 s. The ten-fold classification achieved the highest accuracy of 98.99%, i.e., 0.27 higher than the five-fold classification accuracy and the improved KNN classifier achieved an accuracy of 98.72%. When computation efficiency is a priority, the proposed model's ability to quickly recognize different HR severity levels is significant. [ABSTRACT FROM AUTHOR]

Transformers have demonstrated significant promise for computer vision tasks. Particularly noteworthy is SwinUNETR, a model that employs vision transformers, which has made remarkable advancements in improving the process of segmenting medical images. Nevertheless, the efficacy of training process of SwinUNETR has been constrained by an extended training duration, a limitation primarily attributable to the integration of the attention mechanism within the architecture. In this article, to address this limitation, we introduce a novel framework, called the MetaSwin model. Drawing inspiration from the MetaFormer concept that uses other token mix operations, we propose a transformative modification by substituting attention-based components within SwinUNETR with a straightforward yet impactful spatial pooling operation. Additionally, we incorporate of Squeeze-and-Excitation (SE) blocks after each MetaSwin block of the encoder and into the decoder, which aims at segmentation performance. We evaluate our proposed MetaSwin model on two distinct medical datasets, namely BraTS 2023 and MICCAI 2015 BTCV, and conduct a comprehensive comparison with the two baselines, i.e., SwinUNETR and SwinUNETR+SE models. Our results emphasize the effectiveness of MetaSwin, showcasing its competitive edge against the baselines, utilizing a simple pooling operation and efficient SE blocks. MetaSwin's consistent and superior performance on the BTCV dataset, in comparison to SwinUNETR, is particularly significant. For instance, with a model size of 24, MetaSwin outperforms SwinUNETR's 76.58% Dice score using fewer parameters (15,407,384 vs 15,703,304) and a substantially reduced training time (300 vs 467 mins), achieving an improved Dice score of 79.12%. This research highlights the essential contribution of a simplified transformer framework, incorporating basic elements such as pooling and SE blocks, thus emphasizing their potential to guide the progression of medical segmentation models, without relying on complex attentionbased mechanisms. [ABSTRACT FROM AUTHOR]

A material consequence of climate change is the intensification of extreme precipitation in most regions across the globe. The respective trend signal is already detectable at global to regional scales, but long-term variability still dominates local observational records, which are the basis for extreme precipitation risk assessment. Whether the frequency of extreme events is purely random or subject to a low-frequency internal variability forcing is therefore highly relevant for modelling the expected number of extreme events in a typical observational record. Based on millennial climate simulations, we show that long-term variability is largely random, with no clear indication of low-frequency decadal to multidecadal variability. Nevertheless, extreme precipitation events occur highly irregularly, with potential clustering (11% probability of five or more 100-year events in 250 years) or long disaster gaps with no events (8% probability for no 100-year events in 250 years). Even for decadal precipitation records, a complete absence of any tail events is not unlikely, as, for example, in typical 70-year observational or reanalysis data, the probability is almost 50%. This generally causes return levels – a key metric for infrastructure codes or insurance pricing – to be underestimated. We also evaluate the potential of employing information across neighbouring locations, which substantially improves the estimation of return levels by increasing the robustness against potential adverse effects of long-term internal variability. The irregular occurrence of events makes it challenging to estimate return periods for planning and for extreme event attribution.

Hypertensive retinopathy (HR) results from the microvascular retinal changes triggered by hypertension, which is the most common leading cause of preventable blindness worldwide. Therefore, it is necessary to develop an automated system for HR detection and evaluation using retinal images. We aimed to propose an automated approach to identify and categorize the various degrees of HR severity. A new network called the spatial convolution module (SCM) combines cross-channel and spatial information, and the convolution operations extract helpful features. The present model is evaluated using publicly accessible datasets ODIR, INSPIREVR, and VICAVR. We applied the augmentation to artificially increase the dataset of 1200 fundus images. The different HR severity levels of normal, mild, moderate, severe, and malignant are finally classified with the reduced time when compared to the existing models because in the proposed model, convolutional layers run only once on the input fundus images, which leads to a speedup and reduces the processing time in detecting the abnormalities in the vascular structure. According to the findings, the improved SVM had the highest detection and classification accuracy rate in the vessel classification with an accuracy of 98.99% and completed the task in 160.4 s. The ten-fold classification achieved the highest accuracy of 98.99%, i.e., 0.27 higher than the five-fold classification accuracy and the improved KNN classifier achieved an accuracy of 98.72%. When computation efficiency is a priority, the proposed model’s ability to quickly recognize different HR severity levels is significant.

Log data is an important clue to understanding the behaviour of a system at runtime, but the complexity of software systems in recent years has made the data that engineers need to analyse enormous and difficult to understand. While log-based anomaly detection methods based on deep learning have enabled highly accurate detection, the computational performance required to operate the models has become very high. In this study, we propose an anomaly detection method, SPClassifier, based on sparse features and the internal state of the model, and investigate the feasibility of anomaly detection that can be utilized in environments without computing resources such as GPUs. Benchmark with the latest deep learning models on the BGL dataset shows that the proposed method can achieve competitive accuracy with these methods and has a high level of anomaly detection performance even when the amount of training data is small.

Deep Learning based image quality assessment (IQA) has been shown to greatly improve the quality score prediction accuracy of images with single distortion. However, because these models lack generalizability and the accuracy of multidistortion-based image data is relatively low, designing reliable IQA systems is still an open issue. In this paper, we propose to introduce long-range dependencies between local artifacts and high-order spatial pooling into a convolutional neural network (CNN) model to improve the performance and generalizability of the full-reference IQA (FR-IQA). This long-range dependencies model is based on the hypothesis that local apparent artifacts can affect the overall image quality. The proposed network architecture adopts a non-local means algorithm to establish connections between all positions in the deep feature space and uses the Minkowski function to improve the non-linearity of the spatial pooling. Based on this architecture, a robust FR-IQA system has been constructed and evaluated on three well-known single-distortion-based IQA databases (LIVE, CSIQ, and TID2013) and a multidistortion-based IQA database (MDID). Experimental results demonstrate that, compared with the latest FR-IQA systems, the proposed long-range dependencies-boosted CNN-based FR-IQA system can achieve state-of-the-art performance. A comprehensive cross-database evaluation also shows that the proposed system is sufficiently generalized between different databases and multidistortion-based image data is more useful for training robust image quality metrics.

Delineation of retinal vessels in fundus images is essential for detecting a range of eye disorders. An automated technique for vessel segmentation can assist clinicians and enhance the efficiency of the diagnostic process. Traditional methods fail to extract multiscale information, discard unnecessary information, and delineate thin vessels. In this paper, a novel residual U-Net architecture that incorporates multi-scale feature learning and effective attention is proposed to delineate the retinal vessels precisely. Since drop block regularization performs better than drop out in preventing overfitting, drop block was used in this study. A multi-scale feature learning module was added instead of a skip connection to learn multi-scale features. A novel effective attention block was proposed and integrated with the decoder block to obtain precise spatial and channel information. Experimental findings indicated that the proposed model exhibited outstanding performance in retinal vessel delineation. The sensitivities achieved for DRIVE, STARE, and CHASE_DB datasets were 0.8293, 0.8151 and 0.8084, respectively.

Transformers have demonstrated significant promise for computer vision tasks. Particularly noteworthy is SwinUNETR, a model that employs vision transformers, which has made remarkable advancements in improving the process of segmenting medical images. Nevertheless, the efficacy of training process of SwinUNETR has been constrained by an extended training duration, a limitation primarily attributable to the integration of the attention mechanism within the architecture. In this article, to address this limitation, we introduce a novel framework, called the MetaSwin model. Drawing inspiration from the MetaFormer concept that uses other token mix operations, we propose a transformative modification by substituting attention-based components within SwinUNETR with a straightforward yet impactful spatial pooling operation. Additionally, we incorporate of Squeeze-and-Excitation (SE) blocks after each MetaSwin block of the encoder and into the decoder, which aims at segmentation performance. We evaluate our proposed MetaSwin model on two distinct medical datasets, namely BraTS 2023 and MICCAI 2015 BTCV, and conduct a comprehensive comparison with the two baselines, i.e. , SwinUNETR and SwinUNETR+SE models. Our results emphasize the effectiveness of MetaSwin, showcasing its competitive edge against the baselines, utilizing a simple pooling operation and efficient SE blocks. MetaSwin's consistent and superior performance on the BTCV dataset, in comparison to SwinUNETR, is particularly significant. For instance, with a model size of 24, MetaSwin outperforms SwinUNETR's 76.58% Dice score using fewer parameters (15,407,384 vs 15,703,304) and a substantially reduced training time (300 vs 467 mins), achieving an improved Dice score of 79.12%. This research highlights the essential contribution of a simplified transformer framework, incorporating basic elements such as pooling and SE blocks, thus emphasizing their potential to guide the progression of medical segmentation models, without relying on complex attention-based mechanisms., Competing Interests: The authors declare there are no competing interests., (©2024 Lee and Lee.)

Deep hashing is the mainstream algorithm for large-scale cross-modal retrieval due to its high retrieval speed and low storage capacity, but the problem of reconstruction of modal semantic information is still very challenging. In order to further solve the problem of unsupervised cross-modal retrieval semantic reconstruction, we propose a novel deep semantic-preserving reconstruction hashing (DSPRH). The algorithm combines spatial and channel semantic information, and mines modal semantic information based on adaptive self-encoding and joint semantic reconstruction loss. The main contributions are as follows: (1) We introduce a new spatial pooling network module based on tensor regular-polymorphic decomposition theory to generate rank-1 tensor to capture high-order context semantics, which can assist the backbone network to capture important contextual modal semantic information. (2) Based on optimization perspective, we use global covariance pooling to capture channel semantic information and accelerate network convergence. In feature reconstruction layer, we use two bottlenecks auto-encoding to achieve visual-text modal interaction. (3) In metric learning, we design a new loss function to optimize model parameters, which can preserve the correlation between image modalities and text modalities. The DSPRH algorithm is tested on MIRFlickr-25K and NUS-WIDE. The experimental results show that DSPRH has achieved better performance on retrieval tasks.

Most of existing mesh quality assessment metrics consist of a similar two-stage computation process: first constructing a quality map by comparing the local regions between reference mesh and distorted mesh, then adopting a spatial pooling method to generate the overall quality score. In this paper, we propose a novel spatial pooling method for 3D mesh quality assessment based on percentile weighting strategy. We assign more weight to the severely distorted regions of the mesh at the pooling stage and extend the percentile weighting method by incorporating surface area at the pooling stage. Our analysis indicates that the percentile weighting method has a strong capability to emphasize the local regions with severe distortion of the mesh. We develop a mesh quality metric by pooling the local distances generated by the Tensor-based Perceptual Distance Measure metric with our spatial pooling method. We investigate the influence of the parameters of percentile weighting on the performance, and determine the optimal parameters and unified parameters through empirical tests on three publicly available databases. Experimental results demonstrate the effectiveness of our spatial pooling method and the superiority of our metric over state-of-the-art metrics. [ABSTRACT FROM AUTHOR]

The large number of object categories and many overlapping or closely neighboring objects in large-scale urban scenes pose great challenges in point cloud classification. Most works in deep learning have achieved a great success on regular input representations, but they are hard to be directly applied to classify point clouds due to the irregularity and inhomogeneity of the data. In this paper, a deep neural network with spatial pooling (DNNSP) is proposed to classify large-scale point clouds without rasterization. The DNNSP first obtains the point-based feature descriptors of all points in each point cluster. The distance minimum spanning tree-based pooling is then applied in the point feature representation to describe the spatial information among the points in the point clusters. The max pooling is next employed to aggregate the point-based features into the cluster-based features. To assure the DNNSP is invariant to the point permutation and sizes of the point clusters, the point-based feature representation is determined by the multilayer perception (MLP) and the weight sharing for each point is retained, which means that the weight of each point in the same layer is the same. In this way, the DNNSP can learn the features of points scaled from the entire regions to the centers of the point clusters, which makes the point cluster-based feature representations robust and discriminative. Finally, the cluster-based features are input to another MLP for point cloud classification. We have evaluated qualitatively and quantitatively the proposed method using several airborne laser scanning and terrestrial laser scanning point cloud data sets. The experimental results have demonstrated the effectiveness of our method in improving classification accuracy. [ABSTRACT FROM AUTHOR]

Future extreme precipitation events are expected to be influenced by climate change; however, the robustness of this anthropogenically forced response in respect to projection uncertainty especially for sub-daily extremes is not fully understood. We analyze the impact of anthropogenic climate change on 3 hourly extreme precipitation with return periods ranging between 5 and 50 years over Europe using the RCA4 model ensemble simulations at 0.11° resolution. The robustness of the signals is examined based on a regionalized signal-to-noise (S2N) technique by taking the spatial pooling into account and the efficacy of the regionalization is tested by a sensitivity analysis. The results show an increasing signal in 3 hourly extreme precipitation over Europe for all seasons except summer for which a bipolar pattern (increase in the north and decrease in the south) is discerned. For the business-as-usual scenario RCP8.5, the regionalized winter 3 hourly extreme precipitation signals over 9 × 9 model grid cells are statistically significant in roughly 72%, 65%, 59% and 48% of the European area for 5, 15, 25 and 50 year return periods respectively, while 16%–21% of the area will experience significant changes in summer. The S2N values for 3 hourly extreme precipitation changes rise after the spatial pooling by about a factor of 1.4–1.7 for all seasons except summer when they decline by about a factor of 0.78. The results of sensitivity analysis reveal that the regionalization influence is sensitive—in order of decreasing importance—to season, precipitation time scale, precipitation intensity, emission scenario and model spatial resolution. The precipitation time scale is particularly important seasonally in summer and regionally in south Europe when/where short-duration convective precipitation is dominant.

Many typical applications of object detection operate within a prescribed false-positive range. In this situation the performance of a detector should be assessed on the basis of the area under the ROC curve over that range, rather than over the full curve, as the performance outside the prescribed range is irrelevant. This measure is labelled as the partial area under the ROC curve (pAUC). We propose a novel ensemble learning method which achieves a maximal detection rate at a user-defined range of false positive rates by directly optimizing the partial AUC using structured learning. In addition, in order to achieve high object detection performance, we propose a new approach to extracting low-level visual features based on spatial pooling. Incorporating spatial pooling improves the translational invariance and thus the robustness of the detection process. Experimental results on both synthetic and real-world data sets demonstrate the effectiveness of our approach, and we show that it is possible to train state-of-the-art pedestrian detectors using the proposed structured ensemble learning method with spatially pooled features. The result is the current best reported performance on the Caltech-USA pedestrian detection dataset. [ABSTRACT FROM PUBLISHER]

The output of primary visual cortex (V1) is a piecemeal representation of the visual scene and the response of any one cell cannot unambiguously guide sensorimotor behavior. It remains unsolved how subsequent stages of cortical processing combine (‘pool’) these early visual signals into a coherent representation. We (Webb et al., 2007, 2011) have shown that responses of human observers on a pooling task employing broadband, random dot motion can be accurately predicted by decoding the maximum likelihood direction from a population of motion-sensitive neurons. Whereas Amano et al. (2009) found that the vector average velocity of arrays of narrowband, two-dimensional (2-d) plaids predicts perceived global motion. To reconcile these different results, we designed two experiments in which we used 2-d noise textures moving behind spatially distributed apertures and measured the point of subjective equality between pairs of global noise textures. Textures in the standard stimulus moved rigidly in the same direction, whereas their directions in the comparison stimulus were sampled from a set of probability distributions. Human observers judged which noise texture had a more clockwise global direction. In agreement with Amano and colleagues, observers’ perceived global motion coincided with the vector average stimulus direction. To test if directional ambiguities in local motion signals governed perceived global direction, we manipulated the fidelity of the texture motion within each aperture. A proportion of the apertures contained texture that underwent rigid translation and the remainder contained dynamic (temporally uncorrelated) noise to create locally ambiguous motion. Perceived global motion matched the vector average when the majority of apertures contained rigid motion, but with increasing levels of dynamic noise shifted towards the maximum likelihood direction. A class of population decoders utilizing power-law nonlinearities can accommodate this flexible pooling.

We propose a fast quality metric for depth maps, called fast depth quality metric (FDQM), which efficiently evaluates the impacts of depth map errors on the qualities of synthesized intermediate views in multiview video plus depth applications. In other words, the proposed FDQM assesses view synthesis distortions in the depth map domain, without performing the actual view synthesis. First, we estimate the distortions at pixel positions, which are specified by reference disparities and distorted disparities, respectively. Then, we integrate those pixel-wise distortions into an FDQM score by employing a spatial pooling scheme, which considers occlusion effects and the characteristics of human visual attention. As a benchmark of depth map quality assessment, we perform a subjective evaluation test for intermediate views, which are synthesized from compressed depth maps at various bitrates. We compare the subjective results with objective metric scores. Experimental results demonstrate that the proposed FDQM yields highly correlated scores to the subjective ones. Moreover, FDQM requires at least 10 times less computations than conventional quality metrics, since it does not perform the actual view synthesis. [ABSTRACT FROM PUBLISHER]

We present a new method for automatic classification of indirect immunoflourescence images of HEp-2 cells into different staining pattern classes. Our method is based on a new texture measure called shape index histograms that captures second-order image structure at multiple scales. Moreover, we introduce a spatial decomposition scheme which is radially symmetric and suitable for cell images. The spatial decomposition is performed using donut-shaped pooling regions of varying sizes when gathering histogram contributions. We evaluate our method using both the ICIP 2013 and the ICPR 2012 competition datasets. Our results show that shape index histograms are superior to other popular texture descriptors for HEp-2 cell classification. Moreover, when comparing to other automated systems for HEp-2 cell classification we show that shape index histograms are very competitive; especially considering the relatively low complexity of the method. [ABSTRACT FROM AUTHOR]

The output of primary visual cortex (V1) is a piecemeal representation of the visual scene and the response of any one cell cannot unambiguously guide sensorimotor behavior. It remains unsolved how subsequent stages of cortical processing combine ("pool") these early visual signals into a coherent representation. We (Webb et al., 2007, 2011) have shown that responses of human observers on a pooling task employing broadband, random dot motion can be accurately predicted by decoding the maximum likelihood direction from a population of motion-sensitive neurons. Whereas Amano et al. (2009) found that the vector average velocity of arrays of narrowband, two-dimensional (2-d) plaids predicts perceived global motion. To reconcile these different results, we designed two experiments in which we used 2-d noise textures moving behind spatially distributed apertures and measured the point of subjective equality between pairs of global noise textures. Textures in the standard stimulus moved rigidly in the same direction, whereas their directions in the comparison stimulus were sampled from a set of probability distributions. Human observers judged which noise texture had a more clockwise (CW) global direction. In agreement with Amano and colleagues, observers' perceived global motion coincided with the vector average stimulus direction. To test if directional ambiguities in local motion signals governed perceived global direction, we manipulated the fidelity of the texture motion within each aperture. A proportion of the apertures contained texture that underwent rigid translation and the remainder contained dynamic (temporally uncorrelated) noise to create locally ambiguous motion. Perceived global motion matched the vector average when the majority of apertures contained rigid motion, but with increasing levels of dynamic noise shifted toward the maximum likelihood direction. A class of population decoders utilizing power-law non-linearities can accommodate this flexible pooling. [ABSTRACT FROM AUTHOR]

Abstract: Saccades aimed at spatially extended targets land reliably at central locations determined by pooling information across the target shape [Melcher, D., & Kowler, E. (1999). Shape, surfaces and saccades. Vision Research, 39, 2929–2946; Vishwanath, D., & Kowler, E. (2003). Localization of shapes: Eye movements and perception compared. Vision Research, 43, 1637–1653]. Previous findings of saccadic errors when attempting to look at a target in the midst of distractors encouraged suggestions that pooling occurs indiscriminately, with little or no influence of a selective filter to eliminate the influence of nearby distractors. To determine the effectiveness of filtering, saccadic localization was studied for saccades made to a set of target elements (discs) interleaved with an equivalent set of distractors of a different color. With such interleaved elements, selection and spatial pooling are constrained to occur over the same spatial region. The results showed that filtering was effective and saccadic landing position was determined mainly by the target elements. Concurrent perceptual judgments made about the same stimuli (estimating the mean size of either target or distractor discs) showed better performance for the target discs than distractors, confirming that perceptual attention was allocated to the set of target elements. These results: (1) support the role of attention in setting the input to the spatial pooling process that guides saccades to spatially extended targets, and (2) show that perceptual judgments of mean value, often thought to impose modest attentional demands, are not immune to the constraints of this pre-saccadic filter. [Copyright &y& Elsevier]

Deep hashing is the mainstream algorithm for large-scale cross-modal retrieval due to its high retrieval speed and low storage capacity, but the problem of reconstruction of modal semantic information is still very challenging. In order to further solve the problem of unsupervised cross-modal retrieval semantic reconstruction, we propose a novel deep semantic-preserving reconstruction hashing (DSPRH). The algorithm combines spatial and channel semantic information, and mines modal semantic information based on adaptive self-encoding and joint semantic reconstruction loss. The main contributions are as follows: (1) We introduce a new spatial pooling network module based on tensor regular-polymorphic decomposition theory to generate rank-1 tensor to capture high-order context semantics, which can assist the backbone network to capture important contextual modal semantic information. (2) Based on optimization perspective, we use global covariance pooling to capture channel semantic information and accelerate network convergence. In feature reconstruction layer, we use two bottlenecks auto-encoding to achieve visual-text modal interaction. (3) In metric learning, we design a new loss function to optimize model parameters, which can preserve the correlation between image modalities and text modalities. The DSPRH algorithm is tested on MIRFlickr-25K and NUS-WIDE. The experimental results show that DSPRH has achieved better performance on retrieval tasks. [ABSTRACT FROM AUTHOR]

We propose a new efficient architecture for semantic segmentation, based on a "Waterfall" Atrous Spatial Pooling architecture, that achieves a considerable accuracy increase while decreasing the number of network parameters and memory footprint. The proposed Waterfall architecture leverages the efficiency of progressive filtering in the cascade architecture while maintaining multiscale fields-of-view comparable to spatial pyramid configurations. Additionally, our method does not rely on a postprocessing stage with Conditional Random Fields, which further reduces complexity and required training time. We demonstrate that the Waterfall approach with a ResNet backbone is a robust and efficient architecture for semantic segmentation obtaining state-of-the-art results with significant reduction in the number of parameters for the Pascal VOC dataset and the Cityscapes dataset. [ABSTRACT FROM AUTHOR]

Whereas visual crowding does not greatly affect the detection of the presence of simple visual features, it heavily inhibits combining them into recognizable objects. Still, crowding effects have rarely been directly related to general pattern recognition mechanisms. In this study, pattern recognition mechanisms in visual periphery were probed using a single crowding feature. Observers had to identify the orientation of a rotated T presented briefly in a peripheral location. Adjacent to the target, a single bar was presented. The bar was either horizontal or vertical and located in a random direction from the target. It appears that such a crowding bar has very strong and regular effects on the identification of the target orientation. The observer's responses are determined by approximate relative positions of basic visual features; exact image-based similarity to the target is not important. A version of the "standard model" of object recognition with second-order features explains the main regularities of the data., (© 2014 ARVO.)