Real-Time Blackjack Card Counter (YOLOv8 + CNN)

by Jacob Igo

Overview

This project is a computer vision system that detects and counts playing cards in real time for Blackjack using live video.
It started as a custom CNN-based image classifier and evolved into a YOLOv8 one-stage object detector with integrated tracking and counting logic.

This project is for educational and experimental use only. It is not intended for use in casinos or gambling environments.

---

Goals

• Build a machine learning system capable of identifying and counting playing cards in real time.
• Experiment with different model architectures (custom CNN vs. YOLOv8).
• Integrate object detection, classification, and tracking into a unified CV pipeline.
• Demonstrate end-to-end ML workflow: data preparation → training → deployment.

---

Phase 1: Custom CNN Classifier

Architecture

x = self.pool1(F.relu(self.bn1(self.conv1(x)))) x = self.pool2(F.relu(self.bn2(self.conv2(x)))) x = x.view(-1, 128 * 32 * 32) x = F.relu(self.fc1(x)) x = self.dropout(x) x = self.fc2(x)

The initial approach used a custom convolutional neural network (CNN) trained to classify individual card images by rank (Ace through King).

Dataset

• Dataset: Kaggle Playing Cards Dataset

• Each image contained a single card centered on a plain background.

Training Details

Parameter	Value
Optimizer	Adam
Loss	CrossEntropyLoss
Learning Rate	0.001
Batch Size	32
Epochs	7
Accuracy	~95%
Loss	~0.12

Lessons Learned

• The CNN worked for isolated card images, but failed on real-world video where multiple cards appear at different angles.
• Detection and classification needed to be spatially aware — the CNN alone couldn’t localize multiple cards.
• The project required an object detector rather than just an image classifier.

---

Phase 2: YOLOv8 One-Stage Detector

To address those issues, the project was redesigned using a YOLOv8n (nano) model — a one-stage detector that can detect and classify cards simultaneously.

Dataset Preparation

Each card (e.g. 6d, 6h, kc, ah) was labeled as a separate class. from https://www.kaggle.com/datasets/andy8744/playing-cards-object-detection-dataset

Output Details

Parameter	Value
Precision	%97.8
Recall	%98.9
mAP50	%99
mAP50-95	%95
Loss	%0.7

Phase 3: Card Counting Pipeline

It assigns a numeric rank to each card, basically assigning probability of making the right choice (hit, stand, or double down)

Rank	Count Value
2–6	+1
7–9	0
10–A	-1

When a new tracked card is detected, its rank value is added to the running count.

Technologies Used

Library	Purpose
Supervision	Card Tracking
Numpy	Box Drawing / Coordinate Calculating
OpenCV	Use of Webcam / External Camera / Image Processing