AI-Powered CTR Prediction for Ad Creatives

The ad department launches 50 banners—which one will work? A/B tests require weeks and thousands of impressions. We solve this before launch: an AI system analyzes the visual features of a creative and predicts CTR with an accuracy of ±15% based on historical data. Over 5 years, we have implemented similar solutions for 30+ projects in e-commerce and fintech, guaranteeing a 10–30% CTR increase after optimization. One client, a clothing marketplace, spent a significant budget on creative testing; after implementing the AI system, they cut costs by 40% and increased average CTR by 22%. The system pays for itself in 2–3 months.

How Does AI Predict CTR?

The system extracts two types of features: semantic via CLIP and low-level visual (contrast, saturation, text coverage, presence of faces). We train LightGBM on the log of CTR to normalize the distribution. LightGBM is 3x faster than neural network alternatives with comparable accuracy.

import torch
import torch.nn.functional as F
from transformers import CLIPProcessor, CLIPModel
from PIL import Image
import numpy as np
import cv2

class CreativeFeatureExtractor:
    """
    Multimodal features: CLIP semantic + low-level visual.
    """
    def __init__(self):
        self.clip = CLIPModel.from_pretrained(
            'openai/clip-vit-large-patch14'
        ).eval().cuda()
        self.processor = CLIPProcessor.from_pretrained(
            'openai/clip-vit-large-patch14'
        )

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

    def extract_visual_features(self, image: Image.Image) -> dict:
        """
        Low-level features correlating with CTR:
        - face_area_ratio: face presence (face = +18% CTR per Nielsen)
        - contrast: high contrast → visibility
        - color_harmony: harmonious palette
        - text_coverage: % of image covered by text
        - brightness_variance: visual complexity
        """
        img_array = np.array(image)
        h, w = img_array.shape[:2]

        features = {}

        # Contrast (Michelson)
        gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY)
        features['contrast_michelson'] = float(
            (gray.max() - gray.min()) / (gray.max() + gray.min() + 1e-8)
        )

        # Brightness variance
        features['brightness_variance'] = float(gray.std() / 128.0)

        # Dominant colors (k=5 via k-means)
        pixels = img_array.reshape(-1, 3).astype(np.float32)
        criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 20, 1.0)
        _, labels, centers = cv2.kmeans(
            pixels, 5, None, criteria, 5, cv2.KMEANS_PP_CENTERS
        )
        counts = np.bincount(labels.flatten(), minlength=5)
        dominant_color = centers[counts.argmax()]
        features['dominant_hue'] = float(
            cv2.cvtColor(
                dominant_color.reshape(1,1,3).astype(np.uint8),
                cv2.COLOR_RGB2HSV
            )[0,0,0] / 180.0
        )
        features['dominant_saturation'] = float(dominant_color.max() - dominant_color.min()) / 255.0

        # Face detection
        face_cascade = cv2.CascadeClassifier(
            cv2.data.haarcascades + 'haarcascade_frontalface_default.xml'
        )
        faces = face_cascade.detectMultiScale(gray, 1.1, 4)
        total_face_area = sum(fw * fh for _, _, fw, fh in faces) if len(faces) > 0 else 0
        features['face_area_ratio'] = total_face_area / (h * w)
        features['has_face'] = int(len(faces) > 0)
        features['face_count'] = len(faces)

        # Text region fraction (via morphology)
        _, binary = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
        kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
        gradient = cv2.morphologyEx(binary, cv2.MORPH_GRADIENT, kernel)
        features['text_coverage_estimate'] = float(
            (gradient > 0).mean()
        )

        return features

import pandas as pd
import numpy as np
import lightgbm as lgb
from sklearn.model_selection import cross_val_score
from sklearn.preprocessing import StandardScaler

class CreativeCTRPredictor:
    """
    Train on historical data: (visual_features, clip_embedding) → CTR.
    Target variable: log(CTR) to normalize distribution.
    """
    def __init__(self):
        self.feature_extractor = CreativeFeatureExtractor()
        self.model = lgb.LGBMRegressor(
            n_estimators=500,
            learning_rate=0.05,
            max_depth=6,
            min_child_samples=20,
            subsample=0.8,
            colsample_bytree=0.8,
            reg_lambda=0.1
        )
        self.scaler = StandardScaler()

    def build_feature_vector(self, image: Image.Image) -> np.ndarray:
        visual = self.feature_extractor.extract_visual_features(image)
        clip_emb = self.feature_extractor.extract_clip_features(image)

        visual_vec = np.array(list(visual.values()))
        # CLIP 768-dim + visual features ~10 dim
        return np.concatenate([clip_emb, visual_vec])

    def predict_ctr(
        self, image: Image.Image
    ) -> dict:
        features = self.build_feature_vector(image)
        features_scaled = self.scaler.transform(features.reshape(1, -1))
        log_ctr = self.model.predict(features_scaled)[0]
        ctr_predicted = np.exp(log_ctr)

        # SHAP explainability — top 3 factors
        import shap
        explainer = shap.TreeExplainer(self.model)
        shap_values = explainer.shap_values(features_scaled)

        return {
            'predicted_ctr': round(float(ctr_predicted), 4),
            'ctr_percentile': None,   # filled from train distribution
            'top_factors': self._top_shap_factors(shap_values[0], features)
        }

    def _top_shap_factors(
        self, shap_vals: np.ndarray, feature_vals: np.ndarray, top_k: int = 3
    ) -> list:
        top_indices = np.argsort(np.abs(shap_vals))[::-1][:top_k]
        feature_names = (
            [f'clip_{i}' for i in range(768)] +
            ['contrast', 'brightness_var', 'dominant_hue',
             'dominant_sat', 'face_ratio', 'has_face',
             'face_count', 'text_coverage']
        )
        return [
            {
                'feature': feature_names[i] if i < len(feature_names) else f'feat_{i}',
                'shap_value': float(shap_vals[i]),
                'direction': 'positive' if shap_vals[i] > 0 else 'negative'
            }
            for i in top_indices
        ]

Why LightGBM Instead of DNN?

Gradient boosting on tabular data yields better quality than neural networks with small data volumes (thousands of creatives). LightGBM is 3x faster than CatBoost on 500+ features, and SHAP interpretation works out of the box. For production, we use vLLM for CLIP inference and ONNX Runtime for model acceleration.

What Is SHAP and How Does It Help Designers?

SHAP (SHapley Additive exPlanations) assigns each feature a contribution to the prediction—similar to distributing winnings in a cooperative game. The designer sees that "face presence" increased the predicted CTR by 0.8% and "low contrast" decreased it by 0.4%. This allows conscious banner improvement rather than guessing. We integrate SHAP graphs into a Streamlit dashboard where you can upload a creative and immediately get the top three factors.

How We Integrate the System into Your Creative Workflow

We deploy a REST API on Triton Inference Server or SageMaker. The API accepts an image and returns predicted_ctr and top_factors. Through plugins for Figma and Photoshop, designers get predictions directly in the interface. Built-in MLOps infrastructure (MLflow, DVC) versions data and models, supports A/B tests of new versions. Typical integration takes 1–2 weeks.

Key Insights from Data

Based on accumulated statistics from ad platforms (Google, Meta):

Process of Work

1. Analytics: audit historical creatives, collect metrics, evaluate sample size. We check that there is enough data and balance by category.
2. Design: choose architecture (CLIP vs EfficientNet, LightGBM vs XGBoost), describe feature pipeline. Define quality metrics: MAE, MAPE, Spearman correlation.
3. Implementation: train model, validate on holdout set (time-based split). Integrate SHAP for explainability.
4. Testing: A/B experiment on 20 creatives — compare prediction with actual CTR after a week of impressions.
5. Deployment: deploy API, connect to design pipeline via plugins. Set up data drift monitoring.

What's Included in the Work

• Documentation: model card, operational manual, API description.
• Source code with DVC versioning of data.
• Access to Streamlit dashboard for testing new creatives.
• Team training on the system (2 hours).
• Support for 3 months after deployment.

Timelines

We guarantee prediction transparency: each predicted CTR comes with the top three factors from SHAP. Get a consultation on your project—we'll evaluate your data in 2 days. Contact us to assess your data—we'll analyze 100 creatives in 2 days. Order a pilot on 20 creatives—results in a week.

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.