AI Visual Product Search by Photo

Imagine a customer photographs sneakers on the street and uploads the image to an app — the system finds similar items in the catalog within seconds. No text description, no category selection. This is visual search — a task we solve for e-commerce: from fashion marketplaces to spare parts catalogs. Below is the architecture used in production.

How CLIP Handles Search by Photo

CLIP (OpenAI) produces embeddings of dimension 768 (ViT-L/14) — a common vector space for images and text. This enables not only image-to-image search but also text-to-image: "red Nike sneakers" → similar products. Zero-shot accuracy on a 10K catalog is 74% Recall@10. For specific domains (fashion, furniture, electronics), we perform fine-tuning with learning rate 1e-6, freezing the text encoder. This raises accuracy to 86% (CLIP paper).

Why We Use Qdrant Instead of FAISS

In production scenarios, dynamic filters (price, brand, category) and p99 latency are critical. Qdrant supports combined filters via must conditions during search, whereas FAISS requires rebuilding the index when filters change. On a catalog of 1M products, Qdrant (HNSW) gives 25ms latency — 2–3× faster than alternatives at the same accuracy.

Architecture: Embedding + Vector Search

import torch
import torch.nn.functional as F
from transformers import CLIPProcessor, CLIPModel
from PIL import Image
import numpy as np
import qdrant_client
from qdrant_client.models import Distance, VectorParams, PointStruct

class VisualSearchEngine:
    """
    CLIP-embedding + Qdrant vector DB for search by photo.
    CLIP supports text→image and image→image search out of the box.
    """
    def __init__(
        self,
        qdrant_url: str = 'http://localhost:6333',
        collection_name: str = 'products',
        embedding_dim: int = 768    # CLIP ViT-L/14
    ):
        self.clip_model = CLIPModel.from_pretrained(
            'openai/clip-vit-large-patch14'
        ).eval().cuda()
        self.clip_processor = CLIPProcessor.from_pretrained(
            'openai/clip-vit-large-patch14'
        )

        self.client = qdrant_client.QdrantClient(url=qdrant_url)
        self.collection_name = collection_name
        self._ensure_collection(embedding_dim)

    def _ensure_collection(self, dim: int) -> None:
        if not self.client.collection_exists(self.collection_name):
            self.client.create_collection(
                collection_name=self.collection_name,
                vectors_config=VectorParams(
                    size=dim,
                    distance=Distance.COSINE
                )
            )

    @torch.no_grad()
    def embed_image(self, image: Image.Image) -> np.ndarray:
        inputs = self.clip_processor(
            images=image, return_tensors='pt'
        ).to('cuda')
        emb = self.clip_model.get_image_features(**inputs)
        return F.normalize(emb, dim=-1).cpu().numpy().squeeze()

    @torch.no_grad()
    def embed_text(self, text: str) -> np.ndarray:
        inputs = self.clip_processor(
            text=[text], return_tensors='pt', padding=True
        ).to('cuda')
        emb = self.clip_model.get_text_features(**inputs)
        return F.normalize(emb, dim=-1).cpu().numpy().squeeze()

    def index_product(
        self,
        product_id: str,
        product_image: Image.Image,
        metadata: dict
    ) -> None:
        embedding = self.embed_image(product_image)
        self.client.upsert(
            collection_name=self.collection_name,
            points=[PointStruct(
                id=hash(product_id) % (2**63),
                vector=embedding.tolist(),
                payload={
                    'product_id': product_id,
                    'category': metadata.get('category'),
                    'price': metadata.get('price'),
                    'brand': metadata.get('brand'),
                    **metadata
                }
            )]
        )

    def search_by_image(
        self,
        query_image: Image.Image,
        top_k: int = 20,
        filters: dict = None,       # {'category': 'shoes', 'max_price': 5000}
        score_threshold: float = 0.65
    ) -> list[dict]:
        query_embedding = self.embed_image(query_image)

        # Build Qdrant filters
        qdrant_filter = None
        if filters:
            from qdrant_client.models import Filter, FieldCondition, MatchValue, Range
            conditions = []
            for key, value in filters.items():
                if key == 'max_price':
                    conditions.append(
                        FieldCondition(key='price', range=Range(lte=value))
                    )
                elif key == 'min_price':
                    conditions.append(
                        FieldCondition(key='price', range=Range(gte=value))
                    )
                else:
                    conditions.append(
                        FieldCondition(key=key, match=MatchValue(value=value))
                    )
            qdrant_filter = Filter(must=conditions)

        results = self.client.search(
            collection_name=self.collection_name,
            query_vector=query_embedding.tolist(),
            limit=top_k,
            query_filter=qdrant_filter,
            score_threshold=score_threshold,
            with_payload=True
        )

        return [
            {
                'product_id': r.payload['product_id'],
                'score': round(r.score, 4),
                'metadata': {k: v for k, v in r.payload.items()
                             if k != 'product_id'}
            }
            for r in results
        ]

    def search_by_text(
        self, query_text: str, top_k: int = 20
    ) -> list[dict]:
        """Text-to-image search: 'red Nike sneakers' → results"""
        query_embedding = self.embed_text(query_text)
        results = self.client.search(
            collection_name=self.collection_name,
            query_vector=query_embedding.tolist(),
            limit=top_k,
            with_payload=True
        )
        return [{'product_id': r.payload['product_id'],
                 'score': r.score} for r in results]

Cropping the Object Before Search

from rembg import remove as rembg_remove

def prepare_search_query(
    user_image: Image.Image,
    remove_background: bool = True,
    crop_to_object: bool = True
) -> Image.Image:
    if remove_background:
        # rembg → RGBA
        rgba = rembg_remove(user_image)
        if crop_to_object:
            # Auto-crop to bounding box of opaque pixels
            bbox = rgba.getbbox()   # (left, upper, right, lower)
            if bbox:
                rgba = rgba.crop(bbox)
        # White background behind the object
        background = Image.new('RGB', rgba.size, (255, 255, 255))
        background.paste(rgba, mask=rgba.split()[3])
        return background
    return user_image

Fine-tuning CLIP on Fashion Domain

from transformers import CLIPModel, CLIPProcessor
import torch
from torch.optim import AdamW

def finetune_clip_for_domain(
    model: CLIPModel,
    train_loader,          # (image_tensor, text_tensor) pairs
    num_epochs: int = 10,
    learning_rate: float = 1e-6   # very small LR — CLIP is already well trained
) -> CLIPModel:
    """
    Fine-tuning only the visual encoder.
    Text encoder is frozen — we need it for text→image search.
    """
    for param in model.text_model.parameters():
        param.requires_grad = False

    optimizer = AdamW(
        filter(lambda p: p.requires_grad, model.parameters()),
        lr=learning_rate, weight_decay=0.01
    )

    for epoch in range(num_epochs):
        model.train()
        for batch_images, batch_texts in train_loader:
            outputs = model(
                input_ids=batch_texts['input_ids'].cuda(),
                attention_mask=batch_texts['attention_mask'].cuda(),
                pixel_values=batch_images.cuda()
            )
            # InfoNCE loss
            logits_per_image = outputs.logits_per_image
            labels = torch.arange(
                logits_per_image.shape[0], device='cuda'
            )
            loss = (F.cross_entropy(logits_per_image, labels) +
                    F.cross_entropy(logits_per_image.T, labels)) / 2

            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

    return model

Performance

What's Included

We deliver:

• Indexing pipeline code — in Python, using PyTorch and Qdrant.
• ML model — fine-tuned CLIP (if required) with model card and metrics.
• API service — REST/gRPC endpoints with filter and similarity threshold support.
• Documentation — deployment, configuration, and monitoring guides.
• Team training — 2–3 workshops on pipeline operation and retraining.
• Support — one month post-launch, including bug fixes and latency optimization.

Process: From Request to Production

1. Analytics — we study your catalog: image quality, class distribution, need for fine-tuning.
2. Prototype — in 2–3 weeks we build an MVP on 1,000 items and measure accuracy.
3. Development — implement the production pipeline with Qdrant or pgvector.
4. Testing — load testing (1,000 RPS), validation on nonstandard photos (blur, background).
5. Deployment — deploy in your cloud or on-premises.

Timeline

Budget is determined after an initial assessment of catalog size and customization needs. We evaluate your project in 2 days — contact us for a consultation. With over 5 years of experience in computer vision and more than 10 deployed systems, we guarantee results.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.