When a detection network shrinks its feature maps, small objects in satellite and drone imagery (cars and ships 8 to 30 pixels wide) lose their edges. Recent fixes try to put the lost detail back. The problem is that they cannot tell a real edge from sensor noise, so on grainy, blurred or speckled imagery they may put the noise back too. We propose a downsampling layer that restores lost high-frequency detail only where a classical two-sample statistical test says the neighborhoods on either side of a location really differ. The test is scaled by the local variance, so its threshold rises automatically in noisy regions and falls in clean ones. Because it does this without retraining, the detector should hold up when the noise level at test time differs from the noise level seen in training.
The underlying idea is old. Decades ago, edge detection was recast from "compute a derivative" to "test whether two neighboring pixel populations come from the same distribution." That shift gave edge detectors a built-in, noise-adaptive threshold. What is new here is placing that test inside the learned downsampling step of a modern remote sensing detector, as a nearly parameter-free gate. The claim is specific and testable. At equal accuracy on clean data, a detector with significance-gated downsampling should lose much less accuracy under noise and blur levels it never saw in training than a detector using learned-but-ungated reconstruction or fixed edge filters.
What exists and where the gap is
Three lines of work each supply part of this idea.
The statistical edge detection tradition supplies the mechanism. Boundary And Object Detection In Real World Images uses a Gaussian likelihood-ratio test, a test comparing how well one shared distribution versus two separate ones explains the pixels on each side. Because the pooled local variance appears in its denominator, the threshold scales itself: textured regions demand larger mean shifts before an edge fires. Information-Theoretic Feature Detection in Ultrasound Images handles multiplicative speckle by fitting Rayleigh or Fisher-Tippett distributions to adjacent windows and computing a closed-form symmetric KL divergence. In that paper, derivative-of-Gaussian and Canny detectors "collapsed into a dense web of spurious speckle edges," while the divergence detector did not. The same paper found that parametric models beat nonparametric histogram estimates. That result constrains our design toward closed-form statistics rather than learned density estimates. Active Boundary Loss for Semantic Segmentation shows that KL divergence between adjacent pixels can be differentiated inside a deep network, but only as a training loss on predicted class probabilities, never as part of the feature computation.
The resolution-degradation line motivates the proposal. LIP: Local Importance-Based Pooling replaces fixed pooling with a learned importance-weighted sum and raises small-object AP on COCO from 21.9 to 24.0. A Texture Reconstructive Downsampling for Multi-Scale Object Detection in UAV Remote-Sensing Images (TRD) upsamples the downsampled map, computes the residual against the input, learns a correction and adds it back. This gains 3.2 AP on VisDrone-DET and 3.1 AP on NWPU VHR-10. The research file names the shared weakness plainly: neither operator has an explicit way to distinguish object edges from high-frequency sensor noise, so both risk amplifying it.
The remote-sensing line sets the application regime and the closest competitor. LEGNet targets low-quality overhead imagery with fixed Laplacian-of-Gaussian and Scharr derivative filters in shallow stages and Gaussian smoothing in deep ones. It reaches 80.03% mAP on DOTA-v1.0 with LEGNet-S. Its filters are deterministic derivatives with static widths, and its downsampling is ordinary strided convolution. It therefore inherits exactly the derivative failure the ultrasound work documents, and it has no mechanism for adapting to a new noise level. Hi-ResNet adds an edge-aware distance-transform loss for segmentation. Like the Active Boundary Loss, it acts only through the training objective.
The combination that has not been tried, as far as the file shows, is the one we propose: a variance-normalized hypothesis test computed on features at every downsampling step, used to decide which reconstructed detail survives.
Core Claim
Take TRD's back-projection residual $R$, the high-frequency detail lost by a stride-2 downsampling. Multiply it element-wise by a significance map before adding it back. At each output location and each channel, the significance map splits the local input neighborhood (a 4×4 window) into two halves along four orientations: horizontal, vertical and the two diagonals. For each split it computes a Welch-type statistic and keeps the maximum over orientations:
$$ S = \max_{\theta} \frac{|\mu_A^\theta - \mu_B^\theta|}{\sqrt{(s_A^\theta)^2/n + (s_B^\theta)^2/n + \epsilon}} $$
Here $\mu$ and $s^2$ are the sample mean and variance of the features in half-window $A$ or $B$ under orientation $\theta$, and $n$ is the number of samples per half. The output is the base strided convolution plus $\sigma((S-\tau)/T)\odot R$. In this gate, $\tau$ is a learned threshold per stage and $T$ is a learned temperature per stage. This adds two scalars per stage on top of TRD.
At the stem, the input is raw imagery, where the noise model is known. There, for multiplicative speckle, we replace the Welch statistic with the closed-form Gamma-distribution J-divergence between the two halves, which depends only on the log ratio of their means. This follows the parametric recipe of the ultrasound detector.
The variance denominator is the key ingredient. It is what should let one learned $\tau$ transfer across noise levels, because raising noise inflates $s^2$ and raises the effective threshold automatically. A learned gate calibrated to the training noise would need to be retrained to achieve the same shift.
Research Program
The work has three parts.
First, we build the operator. It drops in wherever TRD would, and it will be implemented with separable box filters so that computing the means and variances over the window costs about as much as a few depthwise convolutions. This addresses the latency concern raised against sliding-window statistics.
Second, we analyze what the gate keeps. On images with ground-truth oriented boxes, we measure the fraction of residual energy that lands within two pixels of object boundaries versus in the background, for ungated TRD and for our gated version, as noise increases. If our hypothesis is right, TRD's boundary fraction falls with noise and ours stays roughly flat.
Third, we test the weak points honestly. Deep features are not Gaussian, which is the parametric-mismatch problem. A 4×4 window gives only 8 samples per half, so the test has low statistical power. We will measure how the statistic behaves on pure-noise inputs at each stage, to check that its null distribution is roughly stable across noise levels. We will also try 6×6 windows to trade localization against power, because the window-scale dilemma is real.
Baselines
The naive reference point is the same detector with plain strided-convolution downsampling, trained and evaluated under the identical protocol. It tells us how much any downsampling change matters at all.
The strongest existing method for the degraded-overhead question is LEGNet. It is the published method built for low-quality remote sensing detection, and it reports 80.03% mAP (LEGNet-S) and 78.96% mAP (LEGNet-T, 3.6M parameters) on DOTA-v1.0, against 75.87% for ResNet-50. For fairness, we will insert our operator into the LEGNet-T backbone's downsampling layers, keeping its edge and Gaussian modules. We then retrain with the same detection head, schedule, augmentations and single-scale protocol, from the same initialization, and report both the reproduced LEGNet-T number and ours. On clean DOTA-v1.0, beating it means matching the reproduced mAP within noise. Under unseen corruption levels, it means a clearly smaller drop.
The strongest downsampling competitor is TRD, which reports +3.2 AP on VisDrone-DET and +3.1 AP on NWPU VHR-10 over YOLOv8. We will reimplement it in the same YOLOv8 setting so that the gain can be compared directly with the published number. We will also place it in the same slots as our operator inside the LEGNet-T backbone. LIP is run in the same slots as a second learned-importance competitor. Every method gets the same number of tuning runs for its hyperparameters.
The ablations isolate the key ingredient. The first ablation removes the variance denominator, so the gate becomes an unnormalized mean difference. The second replaces the statistic with a learned small-convolution gate that has more parameters than ours, which tests whether learning can substitute for the test. The third swaps the Gamma J-divergence at the stem for the Welch statistic, which tests whether the noise model matters.
No reported result was found on Lacuna for any of these methods on noise-corrupted DOTA, DIOR-R or VisDrone, so the corrupted-data numbers for all baselines will come from our own runs.
Why not just…?
Why not just denoise the input first and then use TRD? A speckle or Gaussian denoiser tuned to one noise level would remove much of the noise TRD amplifies. This is a real threat. Its weaknesses are that it has to be re-tuned when the sensor or noise level changes, and that it blurs the very 8-pixel boundaries we want to preserve, which is the window-scale dilemma moved to preprocessing. We will run a denoiser followed by TRD, with the denoiser strength tuned on the training noise level, and evaluate it at unseen levels. If it matches our operator there, the proposal's claim fails.
Why not just let LIP or a learned gate learn noise suppression? A learned importance map can in principle compute something like a local contrast ratio. However, whatever threshold it learns is calibrated to the training noise distribution. The learned-gate ablation, with more capacity than ours and trained on the same data, tests this directly. The prediction is that it ties on in-distribution noise and falls behind at unseen levels. If it does not fall behind, the statistical test adds nothing beyond structure that could be learned.
Why not just use LEGNet's fixed edge filters, or add a derivative branch? The ultrasound comparison in Information-Theoretic Feature Detection in Ultrasound Images is direct evidence that derivative filters produce spurious edges under multiplicative speckle while a distribution test does not. LEGNet's Scharr and LoG widths are also static. Our experiments keep LEGNet's filters and add the gate, so any difference under speckle is attributable to the test.
Why not just train with noise augmentation? This is the most likely strong competitor for in-distribution noise. We will give every baseline the same augmentation over a fixed range of noise levels and evaluate outside that range, both stronger noise and different noise families. That is exactly where a fixed learned calibration should break and a self-scaling threshold should not.
Proposed Experiments
The primary benchmarks are DOTA-v1.0 and DIOR-R for oriented detection and VisDrone-DET and NWPU VHR-10 for small-object UAV detection. DOTA and DIOR are introduced in DOTA: A Large-Scale Dataset for Object Detection in Aerial Images and Object detection in optical remote sensing images: A survey and a new benchmark.
We construct controlled corruption suites that follow the physical models the statistical literature uses. These are additive Gaussian noise, Poisson-Gaussian noise, multiplicative Gamma speckle at 1, 2, 4 and 8 looks (a "look" is an independent radar measurement averaged into a pixel; fewer looks means heavier speckle), and blur by a Gaussian point-spread function. Models are trained on clean data or a narrow noise band and tested across the full grid. The file contains no real SAR detection benchmark, so the speckle results are a physically grounded proxy, not a SAR claim.
We report mAP, small-object AP, and the slope of mAP against noise level. We also report parameters, GFLOPs and FPS, to check that the gate stays within TRD's reported 5% parameter overhead.
The decisive check is this. With clean-data mAP within about 0.5 points of reproduced TRD and LEGNet-T, the gated operator should roughly halve the mAP drop at the two strongest unseen speckle and Poisson-Gaussian levels. In the same conditions, the no-denominator and learned-gate ablations should behave like ungated TRD. The boundary-energy analysis should show why: the residual kept by our gate stays concentrated on object boundaries as noise rises.
If the learned gate or the denoise-then-TRD pipeline closes the gap, we will report that the classical test is not needed. That is itself a useful negative result for the reconstruction-based downsampling line.
References
- Boundary And Object Detection In Real World Images
- Information-Theoretic Feature Detection in Ultrasound Images
- Active Boundary Loss for Semantic Segmentation
- LIP: Local Importance-Based Pooling
- A Texture Reconstructive Downsampling for Multi-Scale Object Detection in UAV Remote-Sensing Images
- LEGNet: Lightweight Edge-Gaussian Driven Network for Low-Quality Remote Sensing Image Object Detection
- Hi-ResNet: Edge Detail Enhancement for High-Resolution Remote Sensing Segmentation
- DOTA: A Large-Scale Dataset for Object Detection in Aerial Images
- Object detection in optical remote sensing images: A survey and a new benchmark