Simplified Object Detection for Manufacturing: Introducing a Low-Resolution Dataset

Description

Abstract

Machine learning (ML), particularly within the domain of computer vision (CV), has established solutions for automated quality classification using visual data in manufacturing processes. Object detection as a CV method for quality classification provides a distinct advantage in enabling the assessment of items within the manufacturing environment regardless of their location in images. However, there are substantial challenges regarding data availability, training examples, and the complexity of incorporating within the subject. Published real-world datasets are too complex in resolution and specific tasks to demonstrate and address the potential for small- and middle-sized enterprises (SMEs) to adopt object detection in their manufacturing processes. In this article, we present a simple 640x640 low-resolution dataset based on plastic bricks for object detection, featuring two quality labels to identify minor surface defects in some instances as an example of quality classification. Analyzing our dataset with a YOLOv5 model on four different dataset sizes, we aim to demonstrate the accuracy of a common object detection model in our simple manufacturing use case, showcasing object detection with low-resolution images and the impact of varying data availability. The mean Average Precision mAP@0.5:0.95 in correctly identifying instances improved from 0.786 to 0.833 as we moved from the smallest data size of 485 instances to the complete dataset of about 1500 instances. While our interest is specifically in showcasing object detection for manufacturing with low-resolution images and limited data availability, the generated data and trained model can serve as a common basis to further investigate object detection tasks on a wider variety of similar quality classification use cases in manufacturing.

Methods

This dataset was exported via roboflow.com on September 14, 2023 at 11:14 AM GMT

Roboflow is an end-to-end computer vision platform that helps you
* collaborate with your team on computer vision projects
* collect & organize images
* understand and search unstructured image data
* annotate, and create datasets
* export, train, and deploy computer vision models
* use active learning to improve your dataset over time

For state of the art Computer Vision training notebooks you can use with this dataset,
visit https://github.com/roboflow/notebooks

To find over 100k other datasets and pre-trained models, visit https://universe.roboflow.com

The dataset includes 1500 images of Lego plastic breaks. The total dataset compromises a size of 85.2 MB.
Lego-bricks are annotated in YOLO v5 PyTorch format. The measured variable (label) are defects on the bricks.

The following pre-processing was applied to each image:
* Auto-orientation of pixel data (with EXIF-orientation stripping)
* Resize to 640x640 (Stretch)

No image augmentation techniques were applied.