Mobile AI Vehicle and VIN Recognition: From TensorFlow to TensorFlow Lite Deployment on Android and iOS

In recent years, the rapid upgrade of mobile hardware and continuous optimization of AI algorithms have made on‑device AI feasible; the 58 Used‑Car mobile team therefore explored vehicle auto‑recognition and VIN‑code recognition on Android and iOS.

Background: Traditional AI services run on backend servers due to limited device compute, causing high traffic, poor user experience, and server overload when images are uploaded for inference.

Pain points: Most apps upload raw photos to the server for vehicle detection, which is inflexible, cumbersome, and consumes bandwidth; the server also bears heavy processing load.

Idea: Run an object‑detection model on the device, filter out irrelevant images locally, and only upload images containing vehicles or VIN codes, thereby reducing traffic and server pressure.

Challenges: Mobile compute limits, model size, and integration of AI libraries on Android/iOS.

AI basics: The team used TensorFlow and its Object_Detection API to train a vehicle detection model, then converted it to TensorFlow Lite (TF‑Lite) for on‑device inference. TF‑Lite offers a smaller footprint and lower latency at the cost of some accuracy.

Vehicle detection project: The workflow included preparing and manually labeling ~1800 vehicle images, splitting them into train/val/test sets, converting XML annotations to CSV and then to TFRecord files, training the model, and visualizing progress with TensorBoard. After several iterations, a quantized MobileNet‑V2 model (~300 KB) was selected.

Data augmentation: Random cropping, brightness, contrast, saturation, and hue adjustments were applied to increase sample diversity, which significantly reduced false‑positive detections.

VIN‑code recognition: Using the same pipeline, ~2889 VIN images (expanded to 34 668 samples with adversarial examples) were processed; the model achieved >90 % confidence with occasional multi‑box errors.

Engineering integration: The TF‑Lite model is downloaded at runtime, keeping the app package small (Android +3.1 MB, iOS OC +1.3 MB, iOS Swift +3.1 MB). The detection flow is: real‑time on‑device coarse classification → upload only relevant images → server performs fine‑grained brand/model classification.

Optimization results: Introducing on‑device detection reduced coarse‑classification latency from ~150 ms to 0 ms and overall response time from ~500 ms to ~350 ms (≈30‑40 % improvement). The app size increase is limited to the dependency library, avoiding a full model bundle.

Demo: Screenshots and GIFs show the real‑time scanning UI and recognition outcomes for both vehicle and VIN detection.

Conclusion & future work: The two AI features demonstrate a complete mobile AI pipeline, providing a template for further on‑device AI scenarios as mobile hardware continues to improve.

References: 1. https://TensorFlow.google.cn/lite

Authors: Li Tong (Backend & Mobile Lead), Ma Zhichao (Senior iOS Engineer), Huang Haijie (Senior Android Engineer) at 58.com Used‑Car Technology Department.