Crowding as a Bottleneck: Reconstructing Normal Images from Pooled Texture Statistics to Expose Geometric Defects

Human peripheral vision has a known failure called crowding. A letter seen out of the corner of the eye is easy to read on its own, but once other letters surround it, the features blend into an unreadable jumble. The best current explanation is that the visual system keeps only statistical summaries of each local region: which orientations are present, how they co-occur, and how they correlate across scales. It does not keep where each feature sits. One consequence is that small shape distortions become invisible in the periphery, because pooling "irons them out" into statistics that look normal. Anomaly detection by reconstruction needs this behavior but usually lacks it. In that approach, a generator trained only on defect-free images is asked to redraw a test image as its nearest normal version, and defects show up in the difference between input and redrawing. The recurring failure is the identity shortcut: a generator expressive enough to draw normal images well also redraws the defect, so the difference vanishes.

We propose to make crowding the bottleneck itself. The generator will not see pixels. It will see only a map of local summary statistics computed over windows of fixed size, and it will be trained on normal data to redraw the image from those statistics. By construction, this input cannot say where things sit inside a window. A geometric defect smaller than the window, such as a bent grid line, a warped tile edge, or a displaced weave, is therefore outvoted by the normal content around it. It is "crowded out," and the generator redraws the normal arrangement. Normal texture, which the statistics describe precisely, is redrawn faithfully. The prediction is sharp and falsifiable. This bottleneck should beat noise-based and resolution-based bottlenecks specifically on anomalies that preserve local color and intensity but break geometry. It should lose on anomalies that change local appearance. Its failure point should follow a size law set by the ratio of defect size to window size.

Core Claim

A fixed, position-discarding pooled-statistics representation is a better bottleneck for reconstruction-based anomaly detection than Gaussian noise or downsampling when the anomaly is geometric. It removes the defect because of what the representation cannot encode, not because of how much capacity the model has or what synthetic corruptions it was shown. It also avoids destroying healthy high-frequency texture, which is the main source of false positives in diffusion reconstructions. Concretely, let $\phi(x)$ be a stack of nonlinear local feature products at each pixel of image $x$: oriented filter magnitudes, their pairwise products across orientation and scale, and local luminance moments, in the spirit of the Portilla–Simoncelli statistics used by the Texture Tiling Model. The bottleneck is

$$ s_w(x)(p) = \sum_q k_w(p - q)\, \phi(x)(q), $$

where $k_w$ is a smooth pooling window of width $w$ centered on location $p$. A decoder $D_w$ is trained on normal images only to minimize a structural-similarity reconstruction loss between $x$ and $D_w(s_w(x))$. The anomaly map sums residuals over a small pyramid of window sizes, $A(x) = \sum_w \lvert x - D_w(s_w(x)) \rvert$, measured in pixel space and in a structural-similarity space.

The size law follows from the pooling. An anomaly occupying a fraction $f$ of a window shifts $s_w$ by roughly $f$ times its statistical difference from the surround. It is suppressed while $f$ is small and passes through once it dominates the window. Detection should therefore degrade sharply as defect area approaches window area. This is a machine analogue of Bouma's law, the rule that crowding spacing scales with pooling-region size.

Where This Sits in Prior Work

The crowding literature supplies both the mechanism and the evidence that it removes geometry. What your visual system sees where you are not looking formulates the Texture Tiling Model, which describes peripheral vision as summary statistics over pooling windows. It shows that images synthesized to match those statistics ("mongrels") predict human peripheral performance. Detecting distortions of peripherally presented letter stimuli under crowded conditions shows that crowding specifically impairs detection of geometric distortions. That is exactly the behavior we want a normal-only reconstructor to have. Both are psychophysics. Neither builds a detector, and we reuse their representation as an engineering component.

Two machine-vision papers constrain the design. Do Deep Neural Networks Suffer from Crowding? identifies spatial pooling as the operation that merges target and flanker features in convolutional networks. It reports that clutter robustness behaves as an architectural property that augmentation does not fix. We invert its framing: pooling that harms recognition is helpful when the "target" is a defect we want erased. COCO-Periph: Bridging the Gap Between Human and Machine Perception in the Periphery applies a uniform version of the Texture Tiling Model across whole images and finds that training detectors on such images improved robustness to geometric corruptions such as elastic transforms. It also reports that generating the dataset took over six months of compute. That cost rules out iterative mongrel synthesis inside a detection loop. For this reason we never synthesize mongrels: we feed the statistics directly to a learned decoder, following the direction of cheap peripheral encoding in Efficient Dataflow Modeling of Peripheral Encoding in the Human Visual System. Optimal visual search based on a model of target detectability in natural images shows that simple blur does not reproduce peripheral feature blending. That result motivates our main comparison, but it is a result about humans; whether it holds for anomaly detection is something we must test, not assume.

On the anomaly side, three lines of work fight the identity shortcut by other means. DRÆM - A discriminatively trained reconstruction embedding for surface anomaly detection trains the reconstructor to remove pasted Perlin-noise texture blobs and adds a segmentation head. Its protection depends on how well synthetic blobs resemble real defects. GLAD: Towards Better Reconstruction with Global and Local Adaptive Diffusion Models for Unsupervised Anomaly Detection adapts the diffusion noise budget per image and fuses reconstructed and original regions. It exists because a fixed amount of Gaussian noise either fails to erase large defects or damages clean texture. DeScarGAN: Disease-Specific Anomaly Detection with Weak Supervision removes the top U-Net skip connection so the generator can repair large deformations, which is a cruder architectural bottleneck. Our proposal differs from all three in one respect. The limit on what can be redrawn comes from a fixed, interpretable representation with a known spatial scale, not from capacity, noise level, or synthetic training data.

Encoding Structure-Texture Relation with P-Net for Anomaly Detection in Retinal Images is the closest in spirit, since it separates structure from texture. It does the opposite of our design, however: it preserves structure through skip connections using a trained segmenter and compresses texture. We preserve texture statistics and discard fine structure, and we need no auxiliary labels. Finally, Learning Global-Local Correspondence with Semantic Bottleneck for Logical Anomaly Detection and Revisiting Deep Feature Reconstruction for Logical and Structural Industrial Anomaly Detection address anomalies that are locally normal but globally wrong. A position-discarding representation is blind to misplacement inside a window, so we treat that setting as a known limitation and probe it rather than claim it.

Research Program

The first component is the bottleneck. We will implement the statistics map $s_w$ as a single feed-forward pass: a steerable filter bank at four scales and four orientations, local products across orientation and adjacent scales, and luminance and color moments. Each is pooled with Gaussian windows at three widths chosen relative to image resolution. This is a deterministic operation costing a few convolutions, so there is no iterative synthesis. We will also build a second instantiation in which $\phi$ is taken from pretrained convolutional features and pooled as means and second moments. This tests whether the benefit comes from the crowding principle (position-free pooled second-order statistics) or from the specific hand-designed statistics.

The second component is the decoder. For each window size, a U-Net-style decoder is trained on normal images only, with structural-similarity plus $\ell_2$ loss. This follows the evidence in MVTec AD - A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection that structural-similarity loss outperforms pure $\ell_2$. The decoder has no pixel skip path, so every piece of information it uses has passed through $s_w$. A diffusion variant conditions a denoiser on $s_w$ and samples from pure noise. This tests whether the bottleneck, not the decoder family, drives the result.

The third component is an analysis of the size law. Using controlled defect insertion on normal images, we vary the defect-to-window area ratio and the defect type. Geometric-only defects are local elastic warps and displaced periodic structure that preserve the local intensity histogram. Appearance defects are color and intensity blobs. We then fit where detection collapses. The crowding account predicts two things. Geometric defects should be detected until they approach window size. Appearance defects should partially leak into the reconstruction at all sizes, because their statistics shift the window's summary. That leakage is the machine counterpart of the feature blending seen in human crowding.

Baselines

The naive reference point is the convolutional autoencoder trained with structural-similarity loss from MVTec AD - A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. It is the plain pixel bottleneck that suffers the identity shortcut. That paper reports that the best model on the Cable category reached an image AUROC of 0.86 but a per-region overlap of only 0.05. That gap between detection and localization is the failure we aim to shrink. We will retrain it with the same decoder architecture as ours, so the only difference is whether the input is pixels or $s_w$. Beating it means higher per-region overlap on every texture category, not only higher image AUROC.

The strongest existing reconstruction method is GLAD, which reports 99.3% image AUROC and 98.6% pixel AUROC on MVTec AD. We do not expect to beat that aggregate, and we will not claim to. The fair comparison is on subsets and failure modes. We will use GLAD's released settings on the same images, evaluate on our geometric-only insertion split and on the geometric defect types within MVTec AD texture categories, and measure false-positive residual energy on defect-free test images. Beating GLAD means higher per-region overlap on geometric defects with lower false-positive residual on normal images, at matched training compute and a matched hyperparameter search. A secondary win condition is that fusing our anomaly map with GLAD's improves on GLAD alone, which would show the two capture complementary information.

DRÆM is the strongest augmentation-based competitor. It reports 98.0 image AUROC and 68.4 pixel average precision on MVTec AD, and 99.0 AUROC on DAGM. We will run it with its Describable Textures Dataset corruption pipeline on identical splits. The test that matters is the geometric-only split, where Perlin texture blobs are a poor stand-in for real defects. There, beating DRÆM means higher pixel average precision.

Two ablations isolate the key ingredient. The first replaces $s_w$ with a downsampled image of equal dimensionality at the same window scale. The second replaces $s_w$ with statistics computed only over the surround of each window, excluding its center, which approximates reconstruction from context. If either matches the full method on geometric defects, the crowding story is not what is doing the work.

Why not just…?

Why not just blur or downsample the input and super-resolve it? This is the most serious threat. A low-resolution bottleneck also discards fine geometry and is much simpler. The human evidence in Optimal visual search based on a model of target detectability in natural images says blur and pooled statistics differ: blur keeps coarse outlines and loses texture, while pooling keeps texture and loses arrangement. For anomaly detection, this predicts that a downsampling bottleneck will redraw fine normal texture poorly, producing false positives on carpet, leather, and wood. It should also retain large low-frequency distortions. The equal-dimensionality downsampling ablation settles this directly. If it matches our method on both false positives and geometric-defect overlap, we will report that crowding adds nothing beyond resolution loss.

Why not just use synthetic anomalies, as DRÆM does, or adaptive noise, as GLAD does? Both work well on aggregate benchmarks, and GLAD's numbers are very high. Their protection against the identity shortcut is statistical, not structural. DRÆM's depends on the corruption distribution. On MVTec LOCO, Learning Global-Local Correspondence with Semantic Bottleneck for Logical Anomaly Detection reports a 24.9% gain in sPRO (a region-overlap localization metric) over DRÆM, showing that DRÆM transfers poorly when real defects differ from its synthetic ones. GLAD's authors built adaptive noise because uniform noise damages clean texture. Our geometric-only split and false-positive measurement test whether a representational bottleneck avoids both problems. If GLAD wins there too, the proposal reduces to an interpretability contribution.

Why not just use pooled deep-feature statistics or a patch memory bank instead of hand-designed texture statistics? Pretrained features may describe texture better than steerable-filter products. On MVTec LOCO, Revisiting Deep Feature Reconstruction for Logical and Structural Industrial Anomaly Detection reports ULSAD at 84.1% image AUROC versus 81.49% for PatchCore, a memory-bank method. We treat this as an instantiation question, not a threat to the claim. That is why the program includes a pooled deep-feature version of $s_w$. The claim is about position-free pooling at a known scale, and either statistic set can confirm or refute it. We will report whichever instantiation performs better.

Why not just mask patches and inpaint them from their surroundings? Reconstruction from context also hides the defect from the generator, and we found no reported result on Lacuna for masked-inpainting anomaly detection on these benchmarks. The difference is that masking removes the normal content under the defect along with the defect. Pooling keeps that content's statistics, which should give more faithful texture redrawing. The surround-only ablation measures this difference.

Proposed Experiments

The main benchmark is MVTec AD, reported separately for the five texture categories and the ten object categories. The texture categories are where pooled statistics are expected to help most, and the object categories are where layout matters and we expect weaker results. We report image AUROC, pixel AUROC, pixel average precision, and per-region overlap. We add DAGM, where DRÆM reports 99.0 AUROC, and MPDD, where GLAD reports a 10.5% pixel-AP gain over earlier reconstruction methods. We also use the synthetic ellipse-deformation dataset from DeScarGAN: Disease-Specific Anomaly Detection with Weak Supervision. Its disease-like shrinking and wave deformation is a nearly pure geometric anomaly. DeScarGAN reports a Dice score of 0.853 there, but under weak supervision, so we cite it as a reference point, not a like-for-like competitor. MVTec LOCO AD is an explicit stress test. We expect missing components to be caught only at the largest window and misplacements within a window to be missed. PatchCore's 81.49% and ULSAD's 84.1% image AUROC are the reference values.

Three checks decide the outcome. First, on the histogram-preserving geometric insertion split, the pooled-statistics bottleneck must beat both the downsampling ablation and GLAD in per-region overlap, at matched false-positive residual on normal images. Second, the size-law experiment must show detection collapsing near a defect-to-window area ratio that shifts proportionally when the window width changes. If the collapse point does not move with window size, the mechanism is not the pooling. Third, the contrast between defect types must hold: our method should gain over baselines on geometric defects and lose ground on appearance defects. If it wins or loses uniformly, the crowding account is wrong even if the numbers are good. Passing all three would establish a scale-controlled, label-free, interpretable bottleneck for reconstruction-based detection. Failing the first or second would mean that texture-statistic pooling offers no advantage over simply throwing away resolution, and we would report it as a negative result.