Custom Image Classifier: From Imbalanced Data to Production

You're deploying a model for a product catalog and get macro-F1 = 0.72 due to severe class imbalance. Recently, an e-commerce project approached us with a dataset of 15,000 images, 30 categories, where 80% belonged to five classes. We applied Weighted Random Sampler and Focal Loss, boosting macro-F1 from 0.72 to 0.94 in two weeks. The key challenge is not the model itself (standard benchmarks are passed), but adaptation to the specific domain: noise in labeling, lighting variations, incomplete data.

We also tackled a medical diagnostic project requiring detection of rare pathologies in MRI scans. The imbalance was even more severe: 99% healthy, 1% diseased. Combining oversampling, augmentation, and focal loss, we achieved sensitivity of 0.92 at specificity of 0.98. Projects start at $3,000 and vary based on complexity.

Which Architecture to Choose?

For most tasks, we select EfficientNet-B4 or ConvNeXt-Tiny: they offer a good balance of accuracy and inference time. EfficientNet-B4 is 2x faster than ViT-B/16 with comparable accuracy. The table below compares popular architectures.

For edge devices (Raspberry Pi, Jetson Nano), we use MobileNetV3 or EfficientNet-Lite – they run in 1–2 ms on CPU.

Why Fine-Tuning Beats Training from Scratch?

Training from scratch requires millions of labeled examples. Fine-tuning a pretrained model yields excellent results with just hundreds of images per class. This approach is described in Wikipedia: Transfer Learning. We also leverage contrastive learning and knowledge distillation to further improve performance.

import timm
import torch.nn as nn

def build_classifier(num_classes: int,
                     pretrained_model: str = 'efficientnet_b4'):
    model = timm.create_model(
        pretrained_model,
        pretrained=True,
        num_classes=0
    )
    embedding_dim = model.num_features  # 1792 for B4

    for param in model.parameters():
        param.requires_grad = False

    classifier = nn.Sequential(
        nn.Linear(embedding_dim, 512),
        nn.GELU(),
        nn.Dropout(0.3),
        nn.Linear(512, num_classes)
    )
    model.classifier = classifier
    return model

Fine-tuning strategy step by step:

1. Freeze backbone, train only classifier for 5 epochs.
2. Unfreeze last 2 blocks, train for 10 epochs with LR 10x lower.
3. Full unfreezing, another 10 epochs with cosine schedule.
4. Evaluate on validation: if metrics are not met, repeat with different hyperparameters.

How to Handle Class Imbalance?

Real datasets are rarely balanced. We combine several techniques:

• Weighted random sampler – sampling frequency inversely proportional to class size.
• Focal Loss – focuses on hard examples (γ=2).
• Oversampling rare classes via data augmentation (albumentations).
• Class-weighted cross-entropy – weights 1/class_frequency.

This approach lifts macro-F1 by 15–20% compared to baseline training. Order a pilot project – we'll show results on your data within two weeks.

Difference Between Multi-Class and Multi-Label Classification

Multi-class: one class per image – softmax + cross-entropy (e.g., animal type). Multi-label: multiple classes simultaneously – sigmoid + binary cross-entropy (e.g., photo tags). The threshold for each class is tuned separately based on F1.

Model Quality Metrics

• Top-1/Top-5 Accuracy for balanced sets.
• Macro-averaged F1 for imbalanced sets.
• Cohen's Kappa for medical tasks.
• AUC-ROC per class for multi-label.

Work Process for Classification System

Analysis → design → implementation → testing → deployment. In the first stage, we study your dataset, identify problematic classes, assess labeling quality. Then we select architecture and run a series of A/B experiments with hyperparameters. After model approval – export to ONNX, containerization, and deployment into your infrastructure. All steps are documented, your engineers receive access to the model and operating instructions. Contact us to get a cost estimate and roadmap for your project.

Timelines

Cost is determined after analysis – contact us for a detailed estimate.

What's Included?

• Data analysis and dataset preparation.
• Architecture selection and fine-tuning (with A/B config tests).
• Quality evaluation per chosen metrics (report).
• Deployment as REST API or integration into your infrastructure.
• Documentation and team training.
• Guarantee – if accuracy does not meet agreed targets, we rework for free.

With years of experience, we have completed over 40 image classification projects for e-commerce, medical, and industrial domains. Our engineers hold certifications from NVIDIA, AWS, and Google Cloud, and use MLOps practices for experiment reproducibility. We guarantee achieving target metrics – if accuracy is below the agreed level, we rework for free. Contact us to discuss your project and get a preliminary timeline estimate.