AI Image Classifier

About AI Image Classifier

How Does Browser-Based Image Classification Work?

Traditional image classification requires sending photos to a cloud API - Google Vision, AWS Rekognition, or Azure Computer Vision all work this way. This tool takes a different approach: the entire neural network runs locally in your browser using WebAssembly (WASM), a binary instruction format that executes at near-native speed. The practical difference is that your images never leave your device.

Step	What Happens
1. Model loading	On first use, ~13 MB of MobileNetV2 weights download via Transformers.js and cache in your browser storage
2. Image preprocessing	Your image is resized to 224 x 224 pixels and normalised to the model's expected RGB input format
3. Inference	ONNX Runtime processes the image through 53 convolutional layers via WebAssembly
4. Classification	The output is a softmax probability distribution over 1,000 ImageNet categories
5. Results	Top 5 predictions are shown ranked by confidence percentage with visual bars

After the initial model download, subsequent classifications typically complete in under a second on modern hardware. The model stays cached in your browser's storage, so repeat visits skip the download entirely.

About MobileNetV2

MobileNetV2 was introduced by Sandler, Howard, Zhu, Zhmoginov, and Chen at Google in their 2018 CVPR paper "MobileNetV2: Inverted Residuals and Linear Bottlenecks." The architecture was specifically designed for mobile and edge devices - small enough to run on a phone but accurate enough for practical classification tasks. Its key innovation is the inverted residual block: instead of compressing data through a bottleneck like traditional residual networks, MobileNetV2 expands the data in a middle layer using lightweight depthwise separable convolutions, then compresses it back down. This keeps the model small while preserving classification quality.

The model was trained on ImageNet-1K (ILSVRC), a benchmark dataset of 1,281,167 training images across 1,000 categories compiled by researchers at Stanford and Princeton (Deng et al., CVPR 2009). ImageNet covers a wide range of everyday objects, animals, vehicles, food, plants, and scenes.

Property	MobileNetV2 1.0
Parameters	~3.5 million
Model size (ONNX quantised)	~13 MB
Input size	224 x 224 pixels, RGB
Output categories	1,000 (ImageNet-1K classes)
Top-1 accuracy on ImageNet	71.8%
Top-5 accuracy on ImageNet	91.0%
Computational cost	~300 million multiply-adds
Architecture	19 inverted residual bottleneck layers
Inference engine	ONNX Runtime via WebAssembly

How Does MobileNetV2 Compare to Other Models?

MobileNetV2 sits at a specific point on the accuracy-vs-size trade-off curve. Larger models like EfficientNet-B0 (5.3M parameters) achieve higher accuracy (~77% top-1) but require more memory and processing time. Smaller variants like MobileNetV3-Small (2.5M parameters) run faster but sacrifice a few percentage points of accuracy. For browser-based inference where download size and execution speed both matter, MobileNetV2 at 3.5M parameters hits a practical sweet spot.

Model	Parameters	Top-1 Accuracy	Best For
MobileNetV2 1.0	~3.5M	71.8%	Browser/mobile inference - good balance of speed and accuracy
MobileNetV3-Large	~5.4M	75.2%	Mobile apps needing slightly better accuracy
MobileNetV3-Small	~2.5M	67.5%	Very constrained devices where speed is critical
EfficientNet-B0	~5.3M	77.1%	Server-side or native apps with more resources
ResNet-50	~25.6M	76.1%	Server-side inference, research benchmarks

Top-1 accuracy means the model's single best guess is correct. Top-5 means the correct answer appears somewhere in the model's five highest-confidence predictions. MobileNetV2's 91.0% top-5 accuracy means it almost always includes the right answer in its results, even when the top guess is wrong.

Understanding Confidence Scores

The confidence percentage next to each prediction represents the model's estimated probability that the image belongs to that category. These probabilities come from a softmax function applied to the network's raw output, which forces all 1,000 category scores to sum to 100%. A high score on the top prediction generally means the model found strong visual evidence for that category. A spread of low scores across multiple predictions usually means the image is ambiguous or unlike anything in the training data.

Confidence Range	What It Means	Typical Scenario
90-100%	Very confident, likely correct	Clear photo of a common object (dog, car, flower)
70-90%	Fairly confident	Good photo, but some ambiguity between similar categories
40-70%	Uncertain	Unusual angle, partial view, or uncommon variant
Below 40%	Low confidence, may be wrong	Image does not match training data well, abstract content

Worked example: If you upload a clear photo of a golden retriever, the model might return "golden retriever" at 87%, "Labrador retriever" at 5%, "cocker spaniel" at 2%, and two other dog breeds at 1% each. The high top score and the fact that all five predictions are dog breeds tells you the model is confident this is a dog - it is just slightly uncertain about the exact breed. By contrast, if the top score is only 15% with unrelated categories, the model is essentially guessing.

What Can the Model Recognise?

MobileNetV2 is trained on ImageNet-1K, which covers 1,000 specific categories. These are heavily weighted toward objects, animals, and scenes that appear frequently in photographs. The model knows 120 dog breeds, dozens of bird species, most common vehicles, many food items, household objects, and outdoor scenes. It does not recognise text, abstract concepts, specific people, logos, or highly specialised technical items outside its training set.

Strong Categories	Limited Categories
Dog breeds (120+ breeds recognised individually)	Abstract art, diagrams, screenshots
Vehicles (cars, trucks, boats, planes, bicycles)	Text in images, documents, handwriting
Food items (pizza, banana, ice cream, broccoli)	Specific people or faces (not trained for this)
Everyday objects (chairs, bottles, phones, clocks)	Logos, brand identification
Nature (flowers, landscapes, insects, mushrooms)	Medical or scientific imagery
Musical instruments (guitar, piano, violin, drum)	Images with multiple dominant subjects

Privacy: Why Local Inference Matters

Cloud-based image classification APIs like Google Vision or AWS Rekognition require uploading your images to external servers. This raises privacy concerns, especially for sensitive photos - personal documents, medical images, location-revealing shots, or proprietary product photos. With browser-based inference via WebAssembly, the image never leaves your device. There is no network request carrying image data, no server-side logging, and no third-party data retention policy to worry about. The ONNX Runtime WebAssembly backend processes everything in your browser's sandboxed environment.

Approach	Data Privacy	Speed	Accuracy	Cost
Cloud API (Google Vision, AWS)	Images sent to external servers	Fast (powerful GPUs)	Very high (large models)	Pay per request
Browser WASM (this tool)	Fully local, nothing uploaded	Moderate (CPU only)	Good (71.8% top-1)	Free, no API key
On-device native app	Fully local	Fast (hardware optimised)	High	App development cost

This makes browser-based classification suitable for use cases where cloud upload is not an option - internal company tools, healthcare triage, or any situation where data sovereignty matters. The trade-off is lower accuracy compared to massive cloud models, but for quick categorisation and prototyping the results are practical.

Practical Uses

Use Case	How It Helps
Developer prototyping	Test image classification UX before committing to a cloud API and its costs
Auto-tagging photos	Quick categorisation for personal or small photo libraries
Machine learning education	See neural network predictions with confidence scores without writing any code
Accessibility exploration	Check what an automated system would describe an image as, useful for alt-text ideas
Content moderation research	Understand how classification models categorise visual content before building production pipelines
Offline capability	After the first model download, classification works without any internet connection

Tips for Better Results

Tip	Why
Use clear, well-lit photos	The model was trained on clean images; dark or blurry photos reduce accuracy
Center the subject	MobileNetV2 expects the main subject roughly centered in the 224x224 frame
Avoid heavy cropping	Very tight crops may remove surrounding context the model relies on
Try different angles	If confidence is low, a different angle of the same object may score higher
Stick to single subjects	Images with one clear subject produce more meaningful results than busy scenes
Check top-5, not just top-1	The correct label often appears in positions 2-5 when the top guess is wrong

Technical Details: Transformers.js and ONNX Runtime

This tool uses Transformers.js by Hugging Face, which provides a JavaScript API for running machine learning models in the browser. Under the hood, Transformers.js loads an ONNX-format version of MobileNetV2 (specifically the Xenova/mobilenet_v2_1.0_224 checkpoint) and executes it through ONNX Runtime's WebAssembly backend. The model weights are quantised to reduce the download from the original ~14 MB float32 weights to roughly 13 MB, with negligible impact on classification accuracy.

The CDN import is pinned to Transformers.js version 3.8.1 to ensure consistent behaviour. After the initial download, the model files are stored in your browser's Cache Storage API, so subsequent visits load almost instantly. Transformers.js v3 also supports a WebGPU backend for GPU-accelerated inference on compatible browsers, though this tool uses the WASM backend for maximum compatibility across devices and operating systems. The WebAssembly approach works on Chrome, Firefox, Safari, and Edge on both desktop and mobile.

For estimating how much memory different AI architectures need, the AI model size calculator breaks down VRAM requirements by model type and precision. The AI token counter handles the text side of AI, counting tokens and estimating API costs across providers. For detecting AI-generated text rather than images, the AI text classifier analyses writing patterns to estimate whether content was human-written or machine-generated.

Sources

• Sandler et al. - MobileNetV2: Inverted Residuals and Linear Bottlenecks (CVPR 2018)
• Deng et al. - ImageNet: A Large-Scale Hierarchical Image Database (CVPR 2009)
• Hugging Face - Transformers.js Documentation
• Google Research - MobileNetV2: The Next Generation of On-Device Computer Vision Networks