LR-YOLO: Lipschitz-Regularized YOLO

• The paper introduces LR-YOLO, which integrates restoration directly into the detector backbone to reduce input and parameter-space sensitivity through Lipschitz regularization.
• LR-YOLO uses a shared backbone and a restoration-aware module to improve feature robustness, addressing instability commonly found in cascaded restoration–detection pipelines.
• Empirical evaluations on hazy and low-light datasets show mAP improvements of up to 2.9, confirming the effectiveness of the integrated, regularized multi-task training strategy.

Searching arXiv for the specified paper and closely related work on restoration–detection integration and YOLO variants. Lipschitz-Regularized YOLO (LR-YOLO) is a YOLO-based detector introduced as the concrete instantiation of the broader Lipschitz-Regularized Object Detection (LROD) framework in "Delving into Cascaded Instability: A Lipschitz Continuity View on Image Restoration and Object Detection Synergy" (Zhao et al., 28 Oct 2025). It is designed for adverse-condition object detection, especially under haze and low light, where image restoration is often used before detection. The method is motivated by the claim that the conventional cascade

$$\text{Image Restoration} \rightarrow \text{Object Detection}$$

is fundamentally unstable because restoration networks and detection networks exhibit different functional behaviors: restoration is smooth and continuous, whereas detection is non-smooth, with sharp decision boundaries. LR-YOLO addresses this mismatch by integrating restoration into shared detector feature learning and imposing Lipschitz-based regularization during training (Zhao et al., 28 Oct 2025).

1. Conceptual basis and problem setting

The motivating problem is the widespread use of restoration-before-detection pipelines for adverse visual conditions such as haze and low light. These pipelines are intuitively appealing because restoration can improve visibility before the detector processes the image. The LROD study argues, however, that such cascades can be unstable because the restoration and detection modules belong to mismatched function classes (Zhao et al., 28 Oct 2025).

The instability is described along two axes. Input-space instability refers to the phenomenon whereby slight perturbations in degradation strength, such as small changes in haze density or luminance, induce large changes in detector features and predictions. Parameter-space instability refers to the difficulty of jointly optimizing restoration and detection through a standard cascade, which leads to rough loss landscapes, sharp gradients, and unstable optimization. The central claim is that the problem is not merely modular composition, but the fact that restoration tends to produce smooth mappings while detection involves discontinuous decision-making around classification and box regression boundaries (Zhao et al., 28 Oct 2025).

This framing places LR-YOLO within a line of work on adverse-condition perception, but its distinctive emphasis is not on stronger restoration alone. Instead, it treats the restoration–detection interface itself as the source of cascaded instability. A plausible implication is that the method should be understood less as a preprocessing refinement and more as a regularized multi-task formulation in which restoration constrains the detector’s representation to become smoother.

2. Lipschitz continuity perspective

The paper analyzes the restoration–detection mismatch through Lipschitz continuity. For a network $f(\cdot;\theta)$, the paper defines $C$-Lipschitz continuity as

$$\|f(\boldsymbol{x}_1;\theta)-f(\boldsymbol{x}_2;\theta)\|_p \le C \|\boldsymbol{x}_1-\boldsymbol{x}_2\|_p .$$

Using the $2$-norm, it recalls that for a differentiable network the Lipschitz constant in input space can be expressed through the Jacobian norm: $$C_x(f) = \sup_{\boldsymbol{x}\in \mathrm{dom}(f)} \|\nabla_{\boldsymbol{x} } f(\boldsymbol{x};\theta)\| .$$ On Pascal VOC with synthetic haze, the empirical observation is that the restoration network has a relatively small Jacobian norm, around 1 to 3.5 per sample, while the object detector has a Jacobian norm nearly an order of magnitude larger (Zhao et al., 28 Oct 2025).

These measurements support the paper’s interpretation that restoration performs smooth, continuous transformation, whereas detection is much more sensitive to small input changes. In a cascade, even if restoration reduces input corruption, residual perturbations can still be amplified by the detector because of its higher Lipschitz constant. The paper summarizes the contrast as follows: restoration is smooth and continuous; detection is discontinuous decision-making; their cascade is unstable because the two behaviors are mismatched (Zhao et al., 28 Oct 2025).

The same perspective is extended to parameter space. The paper defines parameter-space Lipschitz continuity by

$$\|f(\boldsymbol{x};\theta_1)-f(\boldsymbol{x};\theta_2)\| \le C_\theta \|\theta_1-\theta_2\|$$

with constant

$$C_\theta(f)=\sup_{\theta\in\Theta}\|\nabla_\theta f(\boldsymbol{x};\theta)\| .$$

Because parameter updates follow

$$\theta \leftarrow \theta - \mu \nabla_\theta f(\boldsymbol{x};\theta),$$

the parameter-space gradients characterize optimization smoothness. The paper argues that restoration networks exhibit smoother loss landscapes and lower parameter-space Lipschitz constants, while detection networks show rougher surfaces, sharper gradient transitions, and more erratic optimization trajectories. Loss-surface visualizations are reported to show stable convergence for restoration, oscillatory optimization for detection, and disrupted gradient flow in the cascade (Zhao et al., 28 Oct 2025).

3. LROD framework and the definition of LR-YOLO

LROD is the general solution proposed to reduce the mismatch between restoration and detection by aligning their smoothness properties during training (Zhao et al., 28 Oct 2025). LR-YOLO is the YOLO-specific realization of that framework. It is explicitly described not as “YOLO + a separate restoration model,” but as a jointly trained detector whose representation learning is guided by restoration.

The framework operates through two regularization mechanisms. First, input-space regularization via low-Lipschitz restoration embeds restoration directly into detector feature learning rather than treating it as a preprocessing stage. The restoration objective encourages the shared backbone to learn features that are smoother and less sensitive to perturbations. Second, parameter-space regularization via smoothing adds a gradient-norm penalty to stabilize optimization (Zhao et al., 28 Oct 2025).

The practical aim is therefore dual: to reduce detector sensitivity to perturbations in the input and to reduce sensitivity to parameter updates during training. This suggests that LR-YOLO should be interpreted as a regularized co-training regime in which restoration is auxiliary in architecture but central in optimization.

4. Architecture and training objective

LR-YOLO retains the standard YOLO detector decomposition into Backbone, Neck, and Head. It adds a restoration-aware module that taps into the first three backbone stages, uses those low-level features for restoration, and reconstructs a clean image through a restoration-specific head built with CSP-style layers (Zhao et al., 28 Oct 2025). The early backbone features are used because they preserve spatial and textural information useful for restoration while remaining sufficiently early to regularize the shared representation before higher-level detection decisions are formed.

A central architectural property is backbone sharing. The method shares the detector backbone and does not require changing the original YOLO architecture substantially, which is why it is described as a plug-and-play extension for existing YOLO variants (Zhao et al., 28 Oct 2025). The paper states that the framework extends seamlessly to YOLOv8 and YOLOv10.

The total loss is given as

$$\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{det}} + \lambda \cdot \mathcal{L}_{\mathrm{res}} + \lambda_p \cdot \left\|\nabla_{\theta} f_{\theta}(\boldsymbol{x})\right\| ,$$

where $f(\cdot;\theta)$0 is the detection loss, $f(\cdot;\theta)$1 is the restoration loss, $f(\cdot;\theta)$2 is the restoration weight, $f(\cdot;\theta)$3 is the parameter-space regularization weight, and $f(\cdot;\theta)$4 is the gradient-norm regularizer (Zhao et al., 28 Oct 2025). The restoration loss is a Charbonnier loss between restored images and clean ground truth. In the reported experiments, the values are

$f(\cdot;\theta)$5

The ablation study identifies the sharing of the first three stages of the backbone as the best trade-off. According to the paper, shallower sharing gives weak regularization, whereas deeper sharing causes task interference (Zhao et al., 28 Oct 2025). This makes the selected configuration a compromise between smoothness induction and task separation.

5. Theoretical justification

The appendix includes a theorem-like statement intended to justify why restoration regularizes the detector backbone. Let

$f(\cdot;\theta)$6

denote the detector and

$f(\cdot;\theta)$7

denote the restoration branch that shares the same backbone (Zhao et al., 28 Oct 2025).

Under the assumptions that (1) the restoration loss is Lipschitz continuous and its gradient with respect to the backbone is bounded by $f(\cdot;\theta)$8, with

$f(\cdot;\theta)$9

and (2) there exists a training sample $C$0 such that the restoration gradient aligns positively with the gradient of the backbone Jacobian norm, then under continuous-time gradient descent

$C$1

where $C$2 is the unconstrained change induced by the detection loss and $C$3 is the regularizing effect from restoration (Zhao et al., 28 Oct 2025).

The interpretation given in the paper is that the restoration objective pushes the shared backbone toward lower sensitivity, that is, toward a smaller Lipschitz constant, whereas the detection objective may increase sensitivity. With restoration regularization present, the combined effect is a controlled reduction in the backbone Lipschitz constant. This is presented as the theoretical basis for the claim that LR-YOLO is smoother and more stable during training.

6. Benchmarks, empirical behavior, and reported performance

The evaluation covers two adverse-condition settings. For haze, training and validation use synthetic VOC-based haze data: VOC_Haze_Train with 8,111 images and VOC_Haze_Val with 2,734 images; out-of-domain testing uses RTTS with 4,322 real hazy images. Haze is synthesized with the atmospheric scattering model using

$C$4

For low light, training and validation use synthetic VOC-based dark data: VOC_Dark_Train with 12,334 images and VOC_Dark_Val with 3,760 images; out-of-domain testing uses ExDark with 2,563 real low-light images. Low light is simulated with gamma correction using

$C$5

The reported detection metrics are mAP@0.5 and mAP$C$6 for VOC and RTTS, and standard COCO mAP over IoU thresholds $C$7 together with AP$C$8, AP$C$9, AP$$\|f(\boldsymbol{x}_1;\theta)-f(\boldsymbol{x}_2;\theta)\|_p \le C \|\boldsymbol{x}_1-\boldsymbol{x}_2\|_p .$$0, AP$$\|f(\boldsymbol{x}_1;\theta)-f(\boldsymbol{x}_2;\theta)\|_p \le C \|\boldsymbol{x}_1-\boldsymbol{x}_2\|_p .$$1, and AP$$\|f(\boldsymbol{x}_1;\theta)-f(\boldsymbol{x}_2;\theta)\|_p \le C \|\boldsymbol{x}_1-\boldsymbol{x}_2\|_p .$$2 for COCO. Restoration quality is assessed by PSNR and LPIPS (Zhao et al., 28 Oct 2025).

The main reported performance findings are summarized below.

Setting	Reported gain
RTTS, YOLOv10	mAP +2.0
RTTS, YOLOv8	mAP +2.9
ExDark, YOLOv10	mAP +2.9
ExDark, YOLOv8	mAP +2.7

On COCO-based haze and dark settings, the paper reports mAP improvements of about 1.0–1.2 over the compared baselines (Zhao et al., 28 Oct 2025). It also states that LR-YOLO consistently outperforms both cascaded restoration-plus-detection baselines and end-to-end adverse-condition detectors.

Beyond aggregate accuracy, the paper reports that LR-YOLO produces more stable detections under small degradation changes, reduces feature drift in the detector backbone, smooths the loss landscape, and stabilizes optimization trajectories (Zhao et al., 28 Oct 2025). The ablations indicate that restoration loss alone is beneficial, gradient-norm smoothing alone is beneficial, and the combination is best. On RTTS and ExDark, the combination of both regularizers yields the strongest gains over the baseline.

Although the framework is designed for detection, the reported restoration quality is also competitive: strong LPIPS, competitive PSNR, and better restoration quality than the baseline YOLO backbone alone (Zhao et al., 28 Oct 2025). This supports the claim that the restoration branch is not merely an auxiliary loss but a functional component of shared representation learning.

7. Scope, extensions, and interpretation

The paper argues that restoration-plus-detection cascades fail not because restoration is ineffective, but because restoration and detection are optimized in different function classes: restoration is smooth and local, whereas detection is thresholded and discontinuous (Zhao et al., 28 Oct 2025). In a cascade, a smooth module feeds a sensitive, high-Lipschitz detector, so small restoration imperfections may be magnified. LR-YOLO instead allows restoration to shape the detector representation during training, with the stated practical consequences of less sensitivity to degradation fluctuations, more stable optimization, and better accuracy without requiring a separate preprocessing stage. It also reduces inference overhead relative to methods that rely on a full standalone restoration network before detection (Zhao et al., 28 Oct 2025).

The method is presented as especially beneficial for hazy weather detection, low-light object detection, and more generally for adverse-condition recognition tasks in which degradation destabilizes detector behavior. The broader LROD framework is also reported to improve RT-DETR and Faster R-CNN, indicating that the formulation is not tied to YOLO specifically even though LR-YOLO is its showcased implementation (Zhao et al., 28 Oct 2025).

A common misconception would be to treat LR-YOLO as a conventional two-stage pipeline. The paper explicitly rejects that interpretation: restoration is not a separate preprocessing step, but an integrated branch operating on shared early features. Another possible misconception is that the contribution is solely empirical. In fact, the method is structured around an explicit Lipschitz-continuity analysis in both input and parameter space, together with a theorem-like argument linking restoration supervision to reduced backbone sensitivity (Zhao et al., 28 Oct 2025).

In condensed form, LR-YOLO is a Lipschitz-regularized YOLO detector that integrates restoration into shared feature learning and adds parameter-space smoothing in order to resolve the instability produced by the smooth-restoration versus discontinuous-detection mismatch in adverse-condition cascaded pipelines (Zhao et al., 28 Oct 2025).