Development and Validation of an Artificial Neural Network for the Recognition of Custom Dataset with YOLOv4

TL;DR

This paper presents a deep learning-based system using YOLOv4 for real-time recognition of diverse assembly components, demonstrating high accuracy and robustness across various conditions on a Raspberry Pi platform.

Contribution

It develops and validates a novel YOLOv4-based detection system for recognizing complex, varied components in real-time, suitable for embedded systems.

Findings

High recognition accuracy (97-100%) across scenarios

Robust performance unaffected by lighting or orientation

Effective detection of densely packed objects

Abstract

The expanding applications, utilized by more users, enhance hardware performance and further develop cloud systems for big data processing. This leads to numerous unexplored deep learning applications, especially in advanced computer vision for object recognition. Deep learning in image processing encompasses varied tasks from recognizing elements with diverse shapes and sizes to complex element classification, coping with varying backgrounds and lighting conditions, and text recognition. Its advantages lie in robust setup and high performance for recognizing complex elements. This work aims to develop a deep learning-based detection system for automated recognition of assembly components differing in geometry, size, contour, or color. Implementing the YOLOv4 algorithm, the system detects components based on their characteristics. Testing with 13 components involves capturing them in different orientations, numbers, individual parts, or assembled groups using a Raspberry Pi microcontroller and camera. Evaluation focuses on correct object recognition, confidence values, different compositions, distances between objects, and environmental factors affecting system quality. Results show positive object recognition across all scenarios, irrespective of orientation or number of objects. Even densely packed objects are correctly recognized with high confidence (97-100%). Lighting conditions don't significantly impact results, and all objects are properly labeled. The developed system is suitable for real-time two-dimensional component detection, with potential for extension to three-dimensional analysis using multiple cameras with varied positioning and views.