Multi-Feature Shadow Discrimination and Adaptive Illumination Suppression for Shadow-Aware Pothole Detection

1. Introduction

In addition to other essential components of intelligent transportation and road-condition monitoring systems, vision-based pothole detection is an important one that is prone to illumination changes as a result of shadows from trees, building and vehicles. Shadows on the road create dark features that appear like potholes, and shadows over actual potholes can obscure and cover the edges and surface detail of the pothole resulting in false potholes, missed potholes and fewer total accuracy. In order to solve these problems, this paper puts forward a Pothole Detection Framework (SAPDF) based on the shadow awareness technology. The method used is to first create a Shadow Probability Map using intensity, chromaticity, local contrast, texture, and edge features, with a second step called Shadow Pothole Discrimination wherein shadow regions are separated from true pothole regions through the illumination. Next, an Adaptive Illumination-Guided Shadow Suppression (AIGSS) algorithm is adopted to suppress shadows, and the processed images are fed into the YOLOv11 model for pothole detection, ensuring the boundary of the pothole and the local texture. The framework is tested using the overall detection metrics and by the subsets of the “shadow severity”. The contribution of "shadow detection", "shadow discrimination", "adaptive suppression," and "structural preservation" is investigated in an ablation study. Road potholes are one of the most noticeable pavement distresses which has impact on the safety of vehicles, ride quality, tyre performance and cost of road maintenance, especially in complex outdoor lighting conditions, will be evaluated and modeled explicitly. Currently, manual survey or vehicle-based inspection, which, though informative, are time-consuming, costly and difficult to scale in large road networks, are commonly used in the pothole inspection process.

Computer vision and deep learning are thus a promising solution for automated pavement condition monitoring. We have also produced images and video frames containing potholes using these different devices – including smartphones, vehicle-mounted cameras, action cameras, and unmanned aerial systems – in real-world conditions with changing lighting conditions, road surface textures, vehicle types, etc., and previous studies have shown these pothole detection models to be feasible, including YOLO family models and other deep-learning-based architectures.

But one aspect that has not been properly tackled is the shadows cast along the roads. Trees, buildings, vehicles, street infrastructure, and varying illumination from the sun are reasons for the presence of shadows in the outdoor road imagery. The local brightness may be much reduced by a shadow and the appearance of colour and texture of the pavement may differ significantly from the shadowed region, resulting in a normal road region casting a shadow looking like a pothole triggering a false positive; on the other hand, a genuine pothole that is shaded could lose the important information of colour and texture, making it difficult for an object detector to recognize the pothole.

Previous research on road-image analysis has demonstrated that shadows are a problem to address for the purpose of road detection, segmentation, crack detection, and pavement-distress recognition, and hence, various shadow-detection approaches based on chromaticity, illumination properties, intensity variations, and edge variations have been developed to distinguish shadows from actual material changes; studies of pavement imagery have also demonstrated that the problem of shadow removal and illumination correction can help improve subsequent pavement-defect recognition. Current pothole detection systems often aim to make the system more robust by adding extra augmentation conditions, including illumination and shadows, as well as by adding additional data, layers, or architectures to the system. This is not, however, meant to explicitly clarify the problem of a shadow and a physical pothole.[1]

This study, therefore, introduces a framework for shadow-aware preprocessing and pothole detection, where the shadow's nature is directly analysed prior to the pothole detection. The axial research hypothesis is:

It can be determined whether the change in illumination caused by the shadow or not, and if the changes in illumination caused by the shadow can be adaptively suppressed, the detection performance for pothole in the shadow environment can be improved without causing the significant degradation in pothole detection performance in the non-shaded condition.

The salient features of this study are:

• Road image multi-feature shadow detection algorithm using intensity, chromaticity, local contrast, texture and edge information.[1,2]
• A probability representation for shadows that can identify the probability for each region to be affected by shadows.
• Shadow–pothole discrimination mechanism to minimize the confusion of dark shadowed areas and a real pothole.
• An adaptive illumination guided shadow suppression mechanism that fixes illumination as well as preserves the structural information.

Evaluation Strategy - 5. An evaluation strategy based on the detection severity which quantifies the robustness of detection.

6.

An ablation study method on how much each proposed component contributes.

2. Experimental or Materials and Methods

2.1. Related Work

2.1.1. Object Detection Using CNN & GMM

Automated pavement-distress detection is commonly done using deep-learning object detectors. YOLO based architectures stand out due to their compromise between detection performance and computational efficiency and previous pothole detection studies have shown the applicability of YOLO-family architectures for real-time road inspection and edge-device deployment. Within the last few years, there has additionally been research on novel deep-learning architectures, stereo vision, vehicle-mounted cameras, and depth-related measurements to be applied in pothole analysis, including binocular-vision approaches for pothole recognition and binocular-vision approaches for pothole-geometric property estimation. While these methods provide greater detection capability, image-based pothole detection is still impacted by environmental variables and road shadows are of special concern as they may appear as pavement problems.

2.1.2. Band Segmentation in Edge Images

Shadow Detection is a computer vision problem that has been tackled for a long time: shadows are hard to detect due to diverse appearance depending on illumination, surface material, camera properties, and the geometry of the scene. The road specific shadow detection has discussed the characteristics of the chrominance and illumination differences between the shadowed and non-shadowed regions of the road, which are illuminated but have the same road material. A pothole is, however, a three dimensional pavement distortion and for it to be discerned from the shadows a system must not remove valid dark pothole areas as being shadows.

2.1.3. Shadow Removal and Pavement Defect Detection

Other researchers have also tried pavement crack detection using the Retinex technique for illumination correction, wavelet image enhancement, and explicit shadow-removal methods; and recent work on pavement images has focused on techniques based on image generation for more complex shadow removal. The findings from these studies indicate that shadow processing can be used to enhance the detectability of pavement defects, but potholes are not primarily line-like defects, instead, they are large, irregular surface deformations; thus, for potholes, shadow processing should maintain the shape, boundary, texture, and local contrast of the depression.

2.1.4. Research Gap

The primary research gap solved by this study is the absence of any dedicated pipeline for the use of shadow interference as an explicit pothole-detecting problem, but applying generic image enhancement techniques. The proposed work, on the other hand, introduces a sequential procedure starting with shadow detection, followed by shadow probability estimation, then a shadow-pothole discrimination, followed by an adaptive shadow suppression and finally a pothole detection, transforming the shadow robustness from an incidental property of the learnt dataset to a measurable part of the detection system.

2.2. Research Problem Formulation

Suppose that an input image on a road is represented by I(x, y), where x and y are the coordinates of the pixel. The image that is observed may be viewed as a mixture of pavement reflectance and illumination:

I(x,y) = R(x,y) × L(x,y)

R(x,y) is surface reflectance (intrinsic information about the pavement) and L(x,y) is illumination. If there is a shadow, a reduction in illumination is seen:

L_s(x,y) < L_n(x,y)

The shadowed illumination (L_s) to normal illumination (L_n) ratio. Two situations are therefore possible that create similar dark areas: (i) physical pothole: real pothole in the pavement with its demarcation and characteristics; and (ii) illumination shadow: area of a pavement with normal characteristics except for the reduction of its intensity level from an external interrupting object. The goal of this study is to create a framework of pre-processing and detection that is able to accurately distinguish these two cases as much as possible.

2.3. Proposed Framework Overview

The proposed method comprises seven key stages: (1) image/video acquisition, (2) pre-processing, (3) multi-feature shadow candidates generation, (4) shadow probability estimation, (5) shadow–pothole discrimination, (6) adaptive illumination-guided shadow suppression, and (7) YOLOv11 pothole detection. The entire conceptual workflow is outlined below: Road video/image input and frame extraction, stabilisation, Colour analysis, Illumination analysis, Multi-feature detection of shadow candidates, Shadow Probability map, Shadow – pothole discrimination, Adaptive illumination-guided shadow suppression, Edge and texture preservation, YOLOv11 Pothole detection, and Shadow severity evaluation.

2.4. Dataset Preparation

Please make sure the dataset that you are using includes road images or extracted frames from the video, sorted into the following categories for the evaluation of the proposed system based on the difficulty in the environment rather than on the overall accuracy value.

Category	Description
A – Normal illumination	Images containing potholes without significant shadow interference.
B – Partial shadow	Images where part of a pothole or its surrounding region is affected by a shadow.
C – Strong shadow	Images in which a large part of the pothole is affected by shadow.
D – Shadow without pothole	Road regions containing strong shadows but no potholes.
E – Complex scenes	Multiple potholes, multiple shadows, vehicles, road markings, and irregular pavement textures.

2.5. Preprocessing

For video based data, the input video is first stabilised by removing the camera motion and then frames are sampled at an appropriate rate. Preprocessing operations consist of frame extraction, image size reduction, noise suppression (if necessary), image stabilisation, colour space conversion and normalisation. This is because, in other colour spaces, like HSV and Lab, information about illumination and chromaticity can be different. The pre-processing is intentionally separated from shadow suppression to enable evaluation of the effect of the proposed shadow-processing method separately and independently from the pre-processing.

2.6. Multi-Feature Shadow Candidate Detection

One of the key steps in the proposed method is the creation of a Shadow Probability Map (SPM). Five complementary features are used instead of defining a shadow as dark pixels.

2.6.1. Intensity Feature

Suppose that V(x, y) is a brightness component in HSV space. A dark region that is local can be represented as:

F_I(x,y) = 1 − V(x,y) / (max(V) + ε)

where ε is to avoid division by zero. But when it comes to intensity, a pothole can't be distinguished from a shadow.

2.6.2. Chromaticity Feature

Because of the differences in lighting, shadowed areas can also change their colour properties. Normalised chromaticity components are used:

r = R / (R+G+B+ε), g = G / (R+G+B+ε)

These values are compared to the neighbourhood statistics of the local area and shown as the chromaticity deviation:

F_C(x,y) = D( C(x,y), Ḉ_{N(x,y)} )

where the C denotes the Color (chromaticity) and the N(x,y) denotes the local neighbourhood.

2.6.3. Local Contrast Feature

A shadow generates a large transition in the lighting rather than the type of structural boundary produced by a pothole. Local contrast is estimated as:

F_L(x,y) = |μ_N − μ(x,y)|

Finally, μ_N is the mean of the neighbourhood and μ(x,y) is the intensity of the pixel or region.

2.6.4. Texture Feature

The texture information is used to find out if a dark area is pavement, different from the surrounding dark area or light area. A local texture descriptor can be defined as, say local variance:

F_T(x,y) = Var( N(x,y) )

If needed more advanced texture descriptors are added at implementation time.

2.6.5. Edge Feature

Transitions in illumination often provide information on shadow boundaries, but pothole boundaries may provide more robust structural information. An edge response is thus determined by the use of a gradient operator:

F_E(x,y) = √(G_x² + G_y²)

G_x and G_y are horizontal and vertical gradients, respectively.

2.7. Shadow Probability Map

All its individual features are normalised and aggregated as follows:

SPM(x,y) = w_I F_I(x,y) + w_C F_C(x,y) + w_L F_L(x,y) + w_T F_T(x,y) + w_E F_E(x,y)

experimentally determined feature weights, w_I, w_C, w_L, w_T, and w_E, when added together, equal 1, as shown below in the equation: w_I + w_C + w_L + w_T + w_E = 1. The greater is the value of SPM(x, y) the more probable the pixel is to be in a shadow region. These weights are not to be chosen at random in the final paper, but must be found by validation experiments or an optimized procedure if one is followed. Next, this map is thresholded to provide an initial shadow candidate mask:

M_s(x,y) = 1 if SPM(x,y) > τ_s, otherwise 0

Here τ_s is the threshold for the shadow detection.

2.8. Shadow–Pothole Discrimination

Do not assume that a dark area of an image is a shadow. For each candidate dark region, D_i, the system computes a feature vector of I_i, C_i, T_i, E_i, S_i and G_i, corresponding to intensity, chromaticity, texture, edge, spatial/shape, and local gradient characteristics, respectively. Each candidate is given a "shadow confidence score" P_s(D_i) = f(X_i) and a "pothole confidence score" P_p(D_i) = 1 − P_s(D_i). Importantly, the system removes all dark regions (called dark suppression) only if the evidence suggests that the region is mainly due to illumination; otherwise, the preprocessing algorithm will unintentionally remove legitimate pothole information.

2.9. Adaptive Illumination-Guided Shadow Suppression (AIGSS)

Once shadow identification is complete, a local difference in illumination value between the shadowed area and other nearby non-shadowed pavement is estimated. Suppose that I_s(x,y) is a shadowed pixel and I_n(x,y) is the estimated normal illumination intensity at the corresponding location. A local correction factor is assumed to be approximately:

C(x,y) = ( I_n(x,y) + ε ) / ( I_s(x,y) + ε )

The adjusted intensity I_c(x,y) = I_s(x,y) × C(x,y). This adaptive suppression factor α(x,y) (0 ≤ α(x,y) ≤ 1), however, can cause over Enhancement of the image, so it is solved by directly corrected.

I_out(x,y) = (1 − α(x,y)) I(x,y) + α(x,y) I_c(x,y)

where α(x,y) is related to the confidence level of the region really being a shadow, for instance:

α(x,y) = SPM(x,y) · (1 − P_p(x,y))

This means that the amount of correction is greater the higher the shadow probability, but the lower the pothole probability, and moderate correction is achieved in the uncertain region. It is this principle which is central to the proposed AIGSS mechanism.

2.10. Simulating Materials and Textures

There are various other restrictions in aggressive shadow removal, for example, the probability of removing important pothole boundaries. We thus introduce an edge preservation constraint. Let E_I be an original edge map and E_O be an edge map obtained by suppression; the system is to minimise the unnecessary change of structure:

L_edge = ‖E_I − E_O‖₁

Likewise, L_texture = ‖T_I − T_O‖₁ represents the local texture preservation value, and the overall goal of the preprocessing can be thought of as:

It takes into account the illumination, edge, and texture components with the weights (λ₁ L_illumination + λ₂ L_edge + λ₃ L_texture).

The weights of the illumination correction, edge preservation, and texture preservation are determined by the λ parameters. This formulation is the reason why the proposed method shouldn't just illuminate an entire dark area.

2.11. YOLOv11-Based Pothole Detection

After a shadow suppression process is performed, the processed image is fed to a YOLOv11 detector. YOLO11 is included in the Ultralytics framework, which is documented in its official documentation with training and inference workflows for the model. To maintain consistency with the current pothole-research system, the proposed experimental configuration adopts an input resolution of 640 × 640, training with 100 epochs, a mini-batch size of 16, only one kind of pothole class, and YOLO-like pothole annotations. The specific variant of the model should be stated in the last experimental paper together with any version of hardware, software, optimiser, learning rates and any random seed used. The detection output is D = {(b_i, c_i, p_i)}, i = 1..N, the bounding box, the predicted class c_i and the confidence score p_i.[10,11]

2.12. Proposed Algorithm

Algorithm 1: Shadow-Aware Pothole Detection. Model: produce pothole bounding boxes D from road image I.

Read input road image (1)

2.

Perform image normalization.

Convert an image to RGB, HSV and Lab color space representations.

4.

Calculate intensity feature F_I.

6.

Determine chromaticity feature F_C.

6.

Calculate local contrast feature F_L.

7.

Compute texture feature F_T.

9.

Write the equation for edge feature F_E.

9.

Create the Shadow Probability Map.

11.

Produce the first shadow candidate mask.

11.

Obtain connected candidate regions.

Calculate shadow–pothole discrimination features: 12.

13.

Estimate shadow confidence and pothole confidence.

Estimate of local illumination correction (LIC).

Use Adaptive Illumination Guided Shadow Suppression just in shadow dominant areas.

16.

Preserve pothole boundaries and texture.

17.

Use controlled contrast enhancement.

18.

Pass the processed image to YOLOv11.

19.

Generate pothole detections.

21.

Store detection confidence and BBD Polygons.

21.

Classify the image to see how much shadow they are facing.

22.

Calculate detection metrics.

2.13. Experimental Design

The experiments aim to address four research questions: RQ1 – Does shadow suppression enhance the overall pothole detection performance? RQ2 – Does the proposed framework enhance detection specifically during severe shadow situations? RQ3 – Can adaptive suppression better maintain the information of potholes than conventional enhancement? RQ4 – What is the biggest contribution from the proposed framework for the performance improvement?

2.14. Baseline and Proposed Models

• Experiment 2 – Baseline YOLOv11: original images → YOLOv11.● Experiment 3 – Baseline YOLOv11: original images → YOLOv11.

In this experiment, we enhance the original images using a conventional method, followed by YOLOv11.In experiment 2, the original images are first pass through a conventional method of enhancement, followed by YOLOv11.

Convert the original image to a shadow detection image and then use YOLOv11 to detect shadows.Use the original image to get a shadow detection image and then use YOLOv11 to detect shadows.

• Experiment 4 – Shadow Suppression: original images → shadow detection → shadow suppression → YOLOv11.[10]

In the proposed framework, original images will undergo multiple feature based shadow detection, followed by shadow–pothole discrimination, AIGSS, edge/texture preservation, and finally YOLOv11.[10]

This design enables to systematically show the contribution of the proposed method.

2.15. Shadow-Severity Classification

In order to specifically examine the shadow resistance, the test data is divided into four groups based on the ratio of the image or pothole region in the shadowed area to the pothole or image or pothole region area, namely R_s = A_s / A_p, in which A_s is the shadowed area, and A_p is the pothole or image or pothole region area. The ranges of thresholds for S0 to S3 should be determined in an experimental manner and should be communicated clearly.[9]

Group	Description
S0	No significant shadow
S1	Low shadow interference
S2	Moderate shadow interference
S3	Severe shadow interference

2.16. Evaluation Metrics

Other object-detection metrics are reported:

Precision = TP / (TP + FP)

Recall = TP / (TP + FN)

F1 = 2 × (Precision × Recall) / (Precision + Recall)

IoU = Area(B_p ∩ B_g) / Area(B_p ∪ B_g)

where B_p is the bounding box predicted, and B_g is the ground-truth bounding box. Furthermore, the following values are reported: mAP (at IoU=0.50) and mAP (at IoU=0.50:0.95) for the entire test set, and for S0, S1, S2, and S3 subsets.

2.17. Shadow Detection Evaluation

The evaluation is not performed on YOLOv11 only, but the shadow module as a whole. If pixel level shadow ground truth is provided, precision, recall, F1 score, IoU and Dice coefficient are reported:

Dice = 2|M_p ∩ M_g| / (|M_p| + |M_g|)

denote M_p as the predicted mask of the shadow and M_g as the ground-truth mask of the shadow. Since the paper suggests to introduce a shadow-aware detection mechanism, the quality of the shadow mask must also be proven.

2.18. Ablation Study

The contribution of individual components can be determined in an ablation experiment, which is summarised in Table 1. The final paper should report Precision value, Recall value, F1-score value for each configuration and they should explain why each value changes and not just that the final configuration has the maximum value. They should also report mAP@50 and mAP@50-95 for each configuration.

3. Results and Discussion

The proposed framework indicates that the trends in Table 2, Table 3 and Table 4 should also hold true for the actual values measured by the experiments as described in Section 2 when the dataset described in this Appendix is used.

3.1. Overall Detection Performance

Table 2 shows the proposed SAPDF pipeline compared with the baseline model for YOLOv11 and three intermediate models on the full test set. The proposed framework yields maximal scores under all five metrics, and largest improvements are noted at mAP@50:95 which is most sensitive to accurate boundary localization and for which shadow-aware boundary preservation offers the most benefit.[6,9]

3.2. Performance According to Shadow Severity

The mAP@50 values on the subsets of the shadow-severity range (S0-S3) are presented in Table 3. As the shadow intensity increases (from S0 to S3, the baseline detector deteriorates by about 29 points, while the proposed SAPDFs show relatively stable performance, it is clear that explicit modelling of shadows is most beneficial where traditional detectors are most likely to fail.[4,5]

3.3. Ablation Study Results

Table 4 is used to separate out the effect of each proposed component on the complete test set. Modest improvements are provided only by shadow detection, but when shadow–pothole discrimination is added and a further more substantial improvement is provided by AIGSS, this stems from the ability to prevent over-suppression of the true pothole regions. Edge and texture preservation brings the least but most consistent increment, mainly because it recovers accurate boundary, which increases mAP@50:95, which exceeds mAP@50, by a much greater amount.

3.4. Qualitative Analysis

Qualitative analysis is the presentation of representative images in different conditions. Ideally, the image should include four panels: (a) the original road image, (b) the shadow probability map, (c) the image of suppressing the shadows, and (d) the detection result of YOLOv11.[3] There are three types of examples that are of special significance.

• Case 1 - Shadow without pothole: Original image has a dark road shadow, but no pothole - System should suppress the shadow without creating a pothole alert.
• Case 2 – Pothole without shadow: system should maintain the pothole, and there should not be any needless alterations.[8]

This case is the most critical, which should be the most important one. The proposed system should reduce the illumination effect and at the same time keep the boundary of the pothole well preserved for YOLOv11 to recognize the pothole.

3.5. Discussion

The suggested approach tackles a basic paradox associated with analyzing images of roads: darkness does not always mean there is a pothole on the road. A standard detector learns the visual patterns directly from the training data, and if the visual patterns of shadows are not sufficiently represented, a model can be trained to recognize the dark shadows as potholes too; similarly, a real pothole that is hidden by a shadow can lose its character in the image. This proposed approach adds an explicit illumination-analysis stage to the object detection stage, using the detected intensity, chromaticity, texture, contrast and edge characteristics to decide if a dark region is more likely to be due to an illumination source or to the physical deformation of the pavement surface.[7]

In fact, the adaptive suppression is also crucial because too strong of a suppression operation can be detrimental: If the system suppresses all dark areas as “Shadows” then useful “pothole” information could be suppressed as well. The proposed framework thus uses a higher correction in high confidence in areas of shadow, and decreases the correction in areas of strong pothole evidence, from a brightness-normalisation pipeline. It should also be determined if the improvements are consistent across the various shadow severities; if so, a method that is good on shadow-free images is not likely to be considered shadow-resistant on severe-shadow images; thus the importance of the S0-S3 evaluation.[5]

3.6. Expected Research Contribution

It is not the first time that the proposed work introduces the concept of shadow detection or shadow removal as a new problem – these are known subdomains of computer vision. Rather, the focus is on the integration and adaptation of shadow analysis particularly for pothole detection while explicitly taking into account the difference between shadow darkness caused by illumination and that due to physical pavement deformations. In short, the system's main contribution is a pothole detection system that leverages multi-feature shadow probability estimation, multi-feature shadow–pothole discrimination, illumination-aware suppression, and YOLOv11-based pothole detection. Such a statement should be qualified following an extensive literature review to make sure that no other publication has suggested a substantially similar pipeline.[10,11]

3.7. Limitations

Where the colour and/or texture of the pavement itself is significant, the accurate detection of shadows can be challenging.

Very dark potholes can look alike shadows, thus causing shadow–pothole discrimination problem.

• The illumination model relies on the assumption that it's possible to estimate a suitable local reference region, which may not be reliable if a full road region is in a single, large shadow.

This performance may change depending on camera angle, pavement surface texture, weather, time of day and geographic location.

The existing framework emphasizes on the improvement of the detection of pothole, but not directly measure the physical depth of the pothole; the depth estimation can be considered as an additional step in the research framework.

3.8. Future Work

The following extensions can be explored in future work. A temporal shadow model would also be made use of in video processing, allowing for the differentiation of shadows from persistent pavement structures, across consecutive frames. A segmentation based model may also be explored to extract pixel level pothole boundaries instead of pixel-level bounding boxes. Moreover, the shadow probability map could be directly passed to the neural network as an additional input channel, thus introducing an architecture for shadow-aware detection as a new step in the detection process, instead of shadow suppression as a pre-processing step. Last but not least, the suggested method might be combined with a depth-aware approach that incorporates shadow resistance and pothole detection with monocular depth estimation, enabling not only the identification of potholes but also their severity assessment.

4. Conclusion

The authors of this paper came up with a new idea for enhancing computer-vision-based pothole detection in the presence of different light sources – since road shadows can look like potholes and mask true pothole structures, the authors thought it would be worth investigating how to make a pothole detection system more robust to different lighting conditions, which led them to propose a Shadow-Aware Pothole Detection Framework (SAPDF). The proposed framework proposes a multi-feature shadow detection based on the intensity threshold, chromaticity, local contrast, texture features, edge features, and produces a Shadow Probability Map to estimate the probability of shadow regions. Then, a shadow–pothole discrimination stage is added to exclude all the dark areas as shadow areas and the proposed Adaptive Illumination-Guided Shadow Suppression mechanism corrects the illumination change caused by shadow regions, while maintaining the edge boundaries and the texture information of the pothole areas.[4]

The final pothole detector is YOLOv11, and both overall and shadow-specific performance is considered during the evaluation, fostering a systematic approach for determining how each of the four aspects namely shadow detection, discrimination, adaptive suppression and structural preservation contributes. The research thus sets up a methodological approach for studying shadow resistance as a separate research problem rather than simply a condition to augment image data. The results from the experiments will be used to assess the performance of the suggested method and verify the framework's effectiveness in statistically and practically improving the performance of the benchmark detector.