CarSenseAI: A Deep-Learning Pipeline for Real-Time Vehicle Damage Detection, Severity Assessment, and Repair-Cost Estimation

CarSenseAI: A Deep-Learning Pipeline for Real-Time Vehicle Damage Detection, Severity Assessment, and Repair-Cost Estimation

Mithun R

Dept. of CSE

Atria Institute of Technology Bengaluru, India

1AT23CS087

Prof. Pavithra SS

Dept. of CSE Atria Institute of Technology Bengaluru, India

Assistant Professor

Katti Gururaj

Dept. of CSE

Atria Institute of Technology Bengaluru, India

1AT23CS066

Lucky Mahawar

Dept. of CSE

Atria Institute of Technology Bengaluru, India

1AT23CS078

Muhammad Rayyan Zabi

Dept. of CSE, Atria Institute of Technology Bengaluru, India

1AT23CS093

AbstractVehicle damage assessment is traditionally per- formed through manual inspection by insurance surveyors, work- shop technicians, and eet-management personnel, a process that is slow, subjective, and difcult to scale as accident volumes and insurance-claim loads grow. We present CarSenseAI, an end-to-end, camera-based mobile system that automates damage detection, localization, severity classication, and repair-cost esti- mation from ordinary smartphone photographs. We give a formal statement of the assessment task as a composition of four learned maps, and specify each stage by its objective: a YOLOv8 detector trained with a CIoU-based composite loss, a Mask R-CNN segmentation stage supplying pixel-level damage area, a CNN severity classier over a continuous severity score, and a cost regressor whose output is projected onto ontology-derived bounds so that predictions cannot leave the range domain knowledge admits. Synthetic data augmentation via GANs and diffusion models compensates for the scarcity of labelled examples in underrepresented damage categories. The system is delivered as a Flutter mobile application backed by a Node.js/Flask API and a Firebase/MySQL data layer, with GPS-based mechanic discovery, SOS emergency alerts, and service reminders layered on the core pipeline. A review of 41 studies published between 2019 and 2026 shows that most existing systems address only one stage of the assessment workow rather than a complete pipeline; CarSenseAI is designed to close that gap. We describe the architecture, the stage-wise mathematical formulation, the requirements, and the planned evaluation protocol, and outline the open problems that remain: simultaneous multi-damage handling, on-device deployment cost, dataset diversity, and fraud detection.

Index Termsvehicle damage detection, deep learning, YOLOv8, Mask R-CNN, severity classication, repair-cost es- timation, insurance technology, mobile computer vision

1. Introduction

   The automotive industrys rapid expansion has driven a corresponding rise in the number of vehicles on the road, and

   with it a rise in accidents and damage-related incidents. This has created sustained demand for damage-assessment systems that are efcient, reliable, and capable of handling claims at scale. The conventional route, manual inspection by insurance surveyors, workshop technicians, and eet-management per- sonnel, remains widely used but is time-consuming, subjective, and hard to scale when claim volumes are high.

   CarSenseAI is proposed as a modern alternative. It is an AI-driven mobile application that identies, categorizes, and evaluates vehicle damage from photographs captured directly on a smartphone. Using convolutional neural networks to- gether with the YOLOv8 object-detection architecture, the system performs damage analysis with dependable accuracy, recognizing scratches, dents, cracks, and structural distortions; locating the affected regions; grading severity; and producing an estimated repair cost. Beyond the core assessment function, the platform layers on GPS-enabled nearby-mechanic sug- gestions, SOS alerts with live location sharing, maintenance reminders, and general repair guidance.

   The goal is to close the gap between the growing need for fast, unbiased damage assessment and the limitations of manual inspection, via a complete pipeline that converts a raw vehicle photograph into a structured, actionable report without specialized hardware or controlled imaging conditions.

   1. Motivation

      Three forces make this the right moment for an automated pipeline. First, smartphone camera quality has reached a point where ordinary consumer photographs carry enough resolution and dynamic range for ne-grained damage cues, such as hair- line scratches, shallow dents, and small cracks, to be visible without specialized rigs or xed lighting. Second, the object-

      detection and segmentation literature has matured to where real-time, mobile-deployable models (YOLOv8, lightweight Mask R-CNN variants) achieve accuracy that was previously the preserve of server-side, batch-processed pipelines. Third, insurance and eet operators face rising claim volumes and rising expectations of near-instant turnaround, which manual inspection structurally cannot provide: a surveyor visit, a written report, and a manual cost lookup routinely take days, during which a vehicle may sit idle and a claim may stall. Automating the visual-assessment step, without removing a human from the nal approval decision, is therefore both technically feasible and operationally valuable.

   2. Proposed System

   CarSenseAI ingests vehicle images and returns structured damage reports, severity grades, and repair-cost estimates in real time. Unlike systems that address only one stage of the pipeline, it combines YOLOv8-based detection, Mask R-CNN segmentation, CNN-driven severity analysis, and regression- based cost prediction in one framework, targeting insurance claim verication, service workshops, used-vehicle market- places, and eet management. Synthetic data augmentation via diffusion models improves robustness on uncommon damage categories, and the camera-only capture path removes any dependency on specialized sensors.

   The remainder of the paper is organized as follows. Sec- tion II surveys prior work; Section III states the problem and the research gap; Section IV gives the formal, stage-wise formulation of the pipeline; Section V and Section VI describe the architecture and processing methodology; Section VII species requirements; Section IX sets out the evaluation protocol and threats to validity; and Section X concludes.

2. Related Work

   We reviewed 41 studies published between 2019 and 2026, spanning four themes: general damage detection, ne-grained dent and scratch detection, repair-cost estimation, and seg- mentation or transformer-based methods. The 21 works most directly relevant to the design decisions below are cited here.

   1. General vehicle damage detection

      Early work relied on edge detection, texture analysis, and hand-engineered features, which performed acceptably in con- trolled settings but degraded under real-world lighting, view- point, and colour variation. Van Ruitenbeek [7] established an early deep-CNN baseline on more than 5,000 images. Later work layered on transfer learning with ResNet, VGG, and MobileNet, multi-scale YOLO-family detectors, and Mask R- CNN-based instance segmentation. Hasan et al. [2] found AI- based methods consistently outperform manual inspection on speed and consistency, several exceeding 90% classication accuracy; Amodu et al. [3] compared CNN, YOLO, and Mask R-CNN and reported 92% accuracy on a standadized bench- mark, a gure we treat as a modern reference point. Verma et al. [4] paired CNN feature extraction with SVM classication. Mohamad et al. [6] showed that separating detection from

      severity regression improves overall accuracy, a design choice we adopt in (9)(10). Lee et al. [5] introduced a three-quarter- view car-damage dataset and showed that viewpoint-consistent capture protocols materially improve classication accuracy over unconstrained photographs, a nding that shaped our decision to normalize orientation during preprocessing (FR- 01). Tran et al. [1] survey the broader eld and note that most published pipelines are still evaluated on small, single- institution datasets, which limits condence in reported accu- racy gures generalizing to new eets or regions.

   2. Dent- and scratch-specic detection

      Fine surface damage is hard to detect because of subtle appearance and scarce labelled data. Qu et al. [8] used GAN-based augmentation to synthesize scratch patterns for underrepresented classes; Strietzel et al. [9] used ControlNet diffusion models for the same purpose with a measurable gain on real evaluation data, which is the basis for the mixing ratio in (16). Baig et al. [10] built a dedicated dent-detection dataset and showed strong YOLOv8 precision and recall for real-time localization. Zhou et al. [11] target surface-quality inspection in industrial settings and demonstrate that classical computer-vision pipelines remain competitive for narrowly scoped, single-defect-type inspection tasks even as deep learn- ing dominates the general case. Yang et al. [12] extend this line with a broader body-surface-defect classier, reinforcing that dent and scratch detection benets from dedicated, defect- specic model heads rather than a single generic classier.

   3. Repair-cost estimation

      Zhang et al. [13] were among the rst to link ResNet50 damage features to repair-cost ranges via transfer learning. Sailaja et al. [14] unied detection and cost prediction in a single framework to cut end-to-end latency. Ahaggach et al. [15] combined ontology-based reasoning with regression to impose domain constraints on predicted costs, an approach we adopt directly and formalize as the projection in (13). DeepClaim [16] integrated severity analysis and cost estima- tion end to end. Perez-Zarate et al. [17] report that automated cost estimates track certied appraiser estimates closely for common, well-represented damage types but diverge for rare or compound damage, echoing the multi-damage limitation identied in Section III.

   4. Segmentation and transformer-based approaches

      Yusuf et al. [18] applied real-time instance segmentation to identify damaged components individually. Panboonyuen et al. introduced SLICK [19], using selective localization and instance calibration for precise boundary detection, and ALBERT [20], applying bidirectional transformer encoders to part-level segmentation. Khan et al. [21] fused IMU, audio, and camera data via a multi-modal autoencoder to catch minor damage that vision alone misses, illustrating the value of sensor fusion where visual evidence is weak. Across this theme the pattern is consistent: transformer-based segmen- tation delivers the strongest boundary precision reported in

      TABLE I

      Literature survey summary

      With millions of road accidents occurring worldwide each year and insurers processing correspondingly large claim volumes, the demand for an automated, unbiased, and scalable damage- evaluation system is increasingly acute.

      Theme		Representative work	Limitation
      General tection Dent scratch	de- &	Amodu et al. [3]; Verma et al. [4] Qu et al. [8]; Baig et al. [10]	No severity or stage Single damage only	cost type

      B. Problem denition

      The system must accept vehicle photographs and produce a

      Cost estima- tion Segmentation

      / transformer

      Insurance / claims

      Zhang et al. [13]; Ahag- gach et al. [15]

      SLICK [19];

      ALBERT [20]; Yusuf

      et al. [18]

      DeepClaim [16]; Sailaja et al. [14]

      No detection integra- tion

      High compute cost

      Narrow scope, no fraud check

      complete damage-assessment result without a human evaluator in the visual-analysis step. Concretely it must (i) identify and classify the damage type or types present in a single image; (ii) localize the damaged region using bounding boxes and segmentation masks; (iii) grade severity on a consistent, reproducible scale; and (iv) convert the visual assessment into an estimated repair-cost range, across varying image quality and vehicle models, at latency low enough for interactive use. Section IV states this formally.

      Given the survey in Section II, six recurring gaps motivate the design.

      1. Simultaneous multi-damage handling. Most systems as- sume one dominant damage category per image, whereas real vehicles often exhibit several damage types at once.

      2. Incomplete pipelines. Few studies connect detection, severity grading, and cost prediction into one deployable system.

      3. Dataset diversity. Public datasets under-represent vehicle

         Fig. 1. Pipeline-stage coverage of the surveyed themes (Table I). A lled cell means the representative works of that theme address that stage. The bars give the fraction of the four stages covered. No surveyed theme covers all four; closing that gap is CarSenseAIs design target.

         the survey, but at a computational cost that the reviewed papers themselves acknowledge is difcult to reconcile with real-time mobile inference, which is why CarSenseAI treats segmentation as a server-side stage rather than an on-device one (Section IX).

   5. Comparative summary

   Table I summarizes the four themes, their representative methods, and their principal limitation, and Figure 1 renders the same information as stage coverage. The recurring pattern is that most systems solve one stage of the pipeline rather than the whole workow, and that the most accurate segmentation and transformer methods are too compute-heavy for direct mobile deployment.

3. Problem Statement and Research Gap

   A. Problem statement

   Vehicle damage assessment today depends largely on man- ual inspection by insurance surveyors, workshop technicians, and eet-management personnel. This approach is naturally subjective, since different evaluators can reach different con- clusions on the same damage; it is time-consuming; it is location-dependent, since it requires a trained inspector to be physically present; and it is susceptible to both honest human

   models, lighting conditions, viewpoints, and damage categories.

   1. Real-time deployability. The most accurate architectures, particularly transformer-based ones, are too compute- heavy for mobile or edge deployment without signicant optimization.

   2. Fraud detection. Authenticity verication for detected damage is largely absent from existing insurance- oriented systems.

   3. Benchmark standardization. The lack of public, univer- sally accepted benchmarks makes cross-study compari- son unreliable.

   CarSenseAI is scoped to address gaps 14 in its current design; gaps 5 and 6 are agged as open problems (Section IX).

4. Formal Problem Formulation

   Let I RH×W ×3 be a captured image and m = (make, model, year, region) the vehicle metadata supplied by the client. The assessment task is the composition

   F = g fsev fseg fdet, F (I, m) = (D, S, [c, c+]),

   (1)

   where D is the detection et, S the per-region severity grades, and [c, c+] the reported cost interval. Figure 2 shows the stages and the intermediate quantity each one produces.

   A. Detection

   The detector returns n triples

   i=1

   error and, in the insurance context, fraudulent misreporting.

   D = {(bi, yi, pi)}n

   , bi R4, yi Y, pi [0, 1], (2)

   with Y = {scratch, dent, crack, structural}. Boxes are re- gressed with the complete-IoU objective, which adds a centre- distance and an aspect-ratio term to plain IoU so that non-

   directly is trained with class-weighted cross-entropy, because minor damage dominates naturally collected data:

   overlapping predictions still receive gradient:

   Lsev

   = L L c

   yi,c

   log yi,c

   , c

   N

   = , (11)

   3Nc

   i c

   LCIoU = 1 IoU(b, b)+

   2(b, b)

   d2 + v, (3)

   where Nc

   is the number of training regions of class c and N

   2

   h

   h

   where is the centre distance, d the diagonal of the smallest enclosing box, v = 4 ( arctan w arctan w )2 the aspect-

   the total.

   D. Cost estimation under ontology constraints

   ratio penalty, and = v/(1 IoU + v) its adaptive weight.

   The training objective is the usual composite

   craw = g(i, yi, m; ), trained with the Huber loss so that a

   2

   2

   H

   The regressor maps severity and metadata to a raw cost,

   i

   handful of very expensive repairs do not dominate the gradient:

   Ldet = bLCIoU + cLcls + dLDFL, (4)

   1 e2, |e| H,

   raw

   (12)

   with Lcls a binary cross-entropy over classes and LDFL the distribution focal loss on the box-distance distribution. Over-

   L(e) =

   H (|e| 1 H

   ), |e| > , e = c

   c.

   lapping predictions are reduced by non-maximum suppression: box bj is discarded when

   aged over IoU thresholds,

   bk : pk > pj IoU(bj, bk) > nms, nms = 0.45. (5) Detection quality is reported as mean average precision aver-

   The ontology contributes lower and upper bounds £(k, y, m)

   and u(k, y, m) for a given part, damage type, and vehicle, taken from labour-rate and parts tables. The reported estimate is the projection of the raw output onto that interval,

   i

   i

   (13)

   ci = k,y,m(craw) = min( max(craw, £(k, y, m)), u(k, y, m)),

   mAP

   = 1 L 1

   yY

   L AP (t). (6)

   t{0.50,0.55,…,0.95}

   which guarantees a prediction can never fall outside what domain knowledge admits, whatever the regressor extrapolates

   (Figure 4). For an image with several detections, part-level

   n

   50:95

   |Y|

   10

   y

   B. Segmentation and damage extent

   Mask R-CNN is trained with the standard multi-task objec-

   overlap is discounted so that two damages on one panel are not billed twice:

   tive

   Lseg = Lcls + Lbox + Lmask, (7)

   C =

   Li=1

   ci

   k

   L

   L

   k iIk

   ci max ci

   iIk

   , (14)

   where Lmask is the per-pixel binary cross-entropy of the mask branch, applied only to the ground-truth class channel. For detection i lying on vehicle part k, the normalized damage extent is the mask area as a fraction of the part area,

   where Ik indexes detections on part k and k [0, 1] is the shared-labour fraction for that part. The interval nally shown to the user is [C(1 ), C(1 + )] with set from validation residuals. Cost accuracy is reported against appraiser gures as

   i

   i

   N

   N

   a = |Mi| [0, 1], |M | = L 1[M (u, v) = 1]. (8)

   i

   |Pk|

   u,v

   MAE = 1 L 1C C 1, MAPE = 100 L |Ci Ci| .

   Normalizing by part area rather than image area is what makes

   ai comparable across capture distances and vehicle sizes.

   N i i i=1

   N

   i=1

   Ci

   (15)

   C. Severity

   Severity is rst computed as a continuous score combin- ing extent, estimated depth d , and boundary irregularity r

   E. Augmentation and latency budget

   Training data mixes real and synthetic images at a per-class ratio that depends on how under-represented the class is,

   (perimeter over the perimeter of an equal-area disc, a proxy for tearing versus smooth deformation):

   i i N N

   c = min

   max,

   max

   Nc

   c

   , (16)

   i = w1ai + w2di + w3ri,

   3

   L

   wj = 1, wj 0. (9)

   j=1

   so that scarce categories such as cracks receive proportionally more GAN/diffusion samples [8], [9] while common ones are left close to their natural distribution, with max capping the

   The reported grade is the induced three-tier rule

   si =

   moderate, 1 i < 2,

   minor, i < 1,

   severe, i 2,

   synthetic share.

   End-to-end response time is additive over the stages,

   (10)

   T = Tnet + Tpre + Tdet + Tseg + Tsev + Tcost + Tdb, (17)

   and, since the AI layer is stateless per request, c inference workers with service rate under offered load give the

   with 1, 2 calibrated on a validation split rather than xed a priori. Figure 3 plots (9) against a for three depth values and shows the bands (10) induces. The CNN head that predicts si

   response time

   W =

   1 1

   ·

   1

   , =

   < 1, (18)

   c

   which is what xes the worker count needed to hold T under the two-second budget of Section VII at a target claim rate (Figure 6).

5. System Architecture

   TABLE II

   Functional requirements

   ID Description

   FR-01 Capture/upload vehicle images (JPEG, PNG) with automatic

   CarSenseAI follows a modular, layered architecture with four layers: Presentation, API Gateway, Processing, and Data

   FR-02

   lighting and orientation normalization.

   Detect and classify damage types at 85% accuracy on the evaluation set.

   (Figure 5). The layering is deliberate: it lets the mobile client, the AI pipeline, and the persistence layer evolve independently, so the detection model can be retrained and redeployed without a client release, and the client can add UI ows without touching inference code.

   The Presentation Layer is the Flutter mobile application: it captures images, displays assessment reports, and exposes GPS-based mechanic discovery and SOS alerts. The API Gateway Layer, built on Node.js/Flask, routes requests, au- thenticates users, validates input, and formats responses. The Processing Layer hosts the AI pipeline and returns a structured report. The Data Layer, on Firebase/MySQL, stores user data,

   FR-03 Localize damage via bounding boxes (2) and segmentation masks (8).

   FR-04 Classify severity into minor, moderate, severe per (10).

   FR-05 Generate repair-cost ranges under the ontology constraints of (13).

   FR-06 Suggest nearby mechanics via GPS, with contact info, dis- tance, and ratings.

   FR-07 Support SOS alerts with live location sharing to pre- congured contacts.

   FR-08 Send service reminders based on mileage and manufacturer intervals.

   FR-09 Provide basic repair guidance for common damage types.

   TABLE III

   Software stack

   vehicle records, historical assessments, and the reference data

   behind £(·) and u(·) in (13), and drives push notications

   Category Technology Purpose

   through Firebase Cloud Messaging.

   Mobile framework

   Flutter SDK Cross-platform UI

   A. Implementation rationale

   Each choice in Table III follows from a requirement rather

   Language Dart App logic and state Deep learning PyTorch/TensorFlow Training and serving Detection YOLOv8 Real-time damage detec-

   than a default. Flutter was chosen over separate native code- bases because it halves the maitenance surface for a two-

   Segmentation Mask R-CNN Backend API Node.js/Flask

   tion

   Pixel-level boundaries Request orchestration

   platform target while still compiling to native ARM code,

   which matters for camera-preview and image-upload respon- siveness. Node.js/Flask was chosen for the gateway because both have mature middleware for authentication, rate limit- ing, and request validation, which is what the gateway is responsible for. Firebase was chosen for authentication and messaging specically, not as the system of record; MySQL remains the system of record for structured relational data such as historical assessments and cost reference tables, where transactional consistency and query exibility matter more than real-time sync. YOLOv8 and Mask R-CNN were cho- sen over end-to-end transformer detectors because Section II shows transformer-based segmentation, while more accurate on boundary precision, is not yet practical for the latency budget in (17); this is a deliberate accuracy/latency trade-off, not an oversight.

6. Methodology

   Data collection and augmentation. Real-world images are supplemented with GAN- and diffusion-model-generated syn- thetic samples [8], [9] at the per-class ratio of (16), offsetting the scarcity of examples for uncommon damage categories.

   Detection backbone. YOLOv8 performs damage detection and coarse classication under (4), chosen for fast real-time inference and strong small-object performance, consistent with Baig et al. [10].

   Segmentation module. Mask R-CNN produces instance-level masks under (7), supplying the area measure ai of (8) that severity and cost both depend on [18].

   Realtime DB Firebase Auth, cloud messaging

   Relational DB MySQL Persistent storage Location Google Maps API Mechanic discovery Notications FCM Push delivery

   Severity classication. A dedicated CNN implements (9) (11), following the separated detection/severity design of Mo- hamad et al. [6], which improves overall pipeline accuracy relative to a single joint model.

   Cost estimation. A regression module trained under (12), combined with the ontology projection (13) [15], converts severity and vehicle metadata into a bounded estimate, keeping predictions consistent with domain knowledge rather than purely data-driven extrapolation.

   Interface and delivery. The pipeline is exposed via a REST API; the Flutter client uploads images and renders structured reports, severity grades, and cost ranges, alongside the GPS, SOS, and reminder features.

7. Requirements Specification

   1. Functional requirements

      Table II lists the core functional requirements, derived from the problem denition in Section III.

   2. Non-functional requirements

      Interface. The client is built in Flutter for a single An- droid/iOS codebase, using material-design widgets and gesture handling, with an interface designed to minimize the steps between capture and result.

      Fig. 2. Processing pipeline and the intermediate quantity produced by each stage, following the composition in (1).

      Fig. 3. Severity score of (9) against normalized damage area a for three depth values, with w = (0.5, 0.3, 0.2) and, for illustration, r = d. Shaded bands are the decision regions of (10); 1 and 2 are calibrated, not xed.

      Fig. 4. Effect of the ontology projection (13). The raw regressor output is free to extrapolate; the reported estimate is conned to [R(k, y, m), u(k, y, m)], and the shaded band is the interval nally shown to the user. Costs are normalized, not currency values.

      Software stack. Table III lists the technology choices and their role.

      Hardware. Development requires 16 GB RAM, a CUDA- capable GPU with 6 GB VRAM, 256 GB SSD storage, and a multi-core processor. Deployment targets an Android device with 4 GB RAM, a rear camera 8 MP, and Android 8.0+; the backend requires 8 GB RAM and a GPU with 4 GB VRAM per inference worker.

      Latency. The end-to-end budget is T 2 s at the 95th

      Fig. 5. Layered system architecture. The AI Processing Layer implements the maps of (1); it holds no per-request state, which is what permits the horizontal scaling modelled by (18).

      Fig. 6. Worker count required to hold the two-second response budget under (18), for a per-worker service rate of = 4 requests/s. These are design-time budget curves from the queueing model, not measurements.

      percentile, decomposed as in (17); Figure 6 shows the worker count (18) implies for a given claim rate.

   3. Security, privacy, and scalability

   Because CarSenseAI handles vehicle photographs, location data, and, in the insurance-claim case, information tied to a claim record, three cross-cutting concerns shape the design. Data protection: images and location data travel over authen- ticated, encrypted API calls, and the Data Layer separates identity and auth (Firebase) from claim and assessment records (MySQL), limiting the blast radius of any single credential compromise. Access control: the gateway enforces per-user authorization on every request, so a client can retrieve only its own assessment history and cannot enumerate other users records. Scalability: the AI Processing Layer is stateless per request, so inference workers scale horizontally behind the gateway during peak claim periods, for instance after severe weather, with the relation between load, worker count, and response time given by (18).

8. System Model

   The principal use cases, each with its main actor and its pre/postcondition pair: registration and authentication (owner; valid email or phone authenticated session); image capture and upload (owner; session and camera access image queued); damage detection (system; valid image labelled boxes (2)); severity evaluation (system; segmented regions a grade per region (10)); repair-cost prediction (system; grade and metadata bounded estimate (13)); nearby-mechanic discovery (owner; location permission ranked list); SOS alert (owner; congured contacts contacts notied with live location); service-reminder management (owner; mileage on record scheduled notication); and repair-guidance access (owner; completed report guidance for the detected damage type). Each is traceable back to a functional requirement in Table II.

9. Evaluation Plan and Threats to Validity

   1. Planned evaluation methodology

      Because CarSenseAI is at the design-and-implementation stage, this paper species the protocol the system is to be assessed against rather than reporting eld results. Detection and segmentation are to be evaluated with (6) and with mask- level IoU, against a held-out test split stratied by damage type, vehicle model, and lighting condition, so that accuracy is not dominated by whichever category is best represented in training. Severity is to be evaluated with per-class precision, recall, and F1, since the three classes are unlikely to be balanced in naturally collected data; the class weights of (11) address the same imbalance during training. Cost estimation is to be evaluated with (15) against certied appraiser estimates on a shared validation subset, which is how the reviewed literature [13], [17] validates automated cost models against a human baseline. FR-02s 85% accuracy target is the accep- tance threshold for the detection stage before a build is eligible for pilot deployment.

   2. Threats to validity

   Multi-damage handling: the detection stage is not yet vali- dated on images containing several concurrent damage types, and the overlap discount (14) is specied but its k are not yet tted. Deployability: transformer-based segmentation is the most accurate option surveyed but remains too compute-heavy for on-device inference, so CarSenseAI relies on server-side inference. ataset diversity: synthetic augmentation mitigates but does not eliminate the shortage of labelled data. Fraud detection: authenticity verication of submitted images is not yet part of the pipeline and is a priority for the insurance use case. Benchmarking: planned evaluation will use internally curated data rather than a standardized public benchmark, which limits direct comparison with prior systems until such a benchmark exists.

   Planned extensions are a multi-label detection head for concurrent damage types; model compression and quantization to move segmentation on-device; expanded training data across makes and imaging conditions; and an authenticity-verication module for fraud-resistant insurance workows.

10. Conclusion

CarSenseAI is a complete, deployable pipeline for vehi- cle damage assessment that composes YOLOv8 detection, Mask R-CNN segmentation, CNN severity classication, and ontology-constrained regression cost estimation (1), wrapped in a Flutter client with GPS, SOS, and reminder features. Against a survey of 41 studies from 20192026, most existing systems address a single stage of this workow; the con- tribution here is to integrate detection, localization, severity, and cost into one production-oriented, camera-only system with an explicit stage-wise formulation, in which the ontology projection (13) keeps cost predictions inside limits domain knowledge admits and the queueing relation (18) ties the accu- racy/latency trade-off to a concrete deployment budget. Con- current multi-damage detection, on-device deployment cost, dataset diversity, and fraud detection dene the immediate next steps.

References

1. D. H. Tran et al., A survey on vehicle damage detection using deep learning, 2025.

2. M. Hasan et al., Vehicle damage detection using articial intelligence: a systematic literature review, 2025.

3. O. D. Amodu et al., A comprehensive evaluation of deep learning models for vehicle damage detection, 2025.

4. N. Verma et al., Car damage detection analysis using deep learning and computer vision techniques, 2025.

5. D. Lee et al., Automated vehicle damage classication using the three- quarter view car damage dataset and deep learning approaches, 2024.

6. A. T. Mohamad et al., Car damage severity assessment using supervised deep learning, 2024.

7. R. E. van Ruitenbeek, Vehicle damage detection using deep convolu- tional neural networks, 2019.

8. G. Qu et al., Automotive scratch detection: a lightweight convolutional network approach augmented by generative adversarial learning, 2025.

9. J. Strietzel et al., AI scratching your car: using diffusion models for training data generation in automotive damage detection, 2024.

10. D. Z. Baig et al., Small dents, big impact: a dataset and deep learning approach for vehicle dent detection, 2025.

11. Q. Zhou et al., An automatic surface defect inspection system for automobiles, 2019.

12. Z. Yang et al., Research on detection and classication of automotive body surface defects, 2025.

13. M. Zhang et al., AI-based vehicle damage repair price estimation system, 2023.

14. N. V. Sailaja et al., Enhanced deep learning techniques for intelligent vehicle damage detection and cost estimation, 2025.

15. H. Ahaggach et al., Enhancing car damage repair cost prediction: integrating ontology reasoning with regression models, 2024.

16. An AI-driven framework for automated vehicle damage severity assess- ment and insurance cost estimation (DeepClaim), 2025.

17. S. A. Perez-Zarate et al., Automated car damage assessment using computer vision, 2024.

18. S. A. Yusuf et al., Automotive parts assessment: applying real-time instance segmentation for automated vehicle damage assessment, 2022.

19. T. Panboonyuen et al., SLICK: selective localization and instance calibration for car damage segmentation, 2025.

20. T. Panboonyuen et al., ALBERT: advanced localization and bidirec- tional encoder representations for transformer-based car damage seg- mentation, 2025.

21. S. Khan et al., Robust anomaly detection through multi-modal autoen- coder fusion for small vehicle damage detection, 2025.