ARG_FIXINGS2 Computer Vision Project

Project Description: RoboFlow Object Detection for Hardware Components

The objective of this project is to develop an object detection model using RoboFlow to accurately identify four common hardware components: bolts, nuts, washers, and screws. This will facilitate the automation of tasks related to component identification and sorting in various industrial applications.

Dataset Preparation: The dataset comprises 203 images in total, with an average image size of 2.07 MP and a resolution of 1920x1080. Out of these, 103 images (batch1) have been manually collected by myself and annotated with 3,440 annotations, yielding a median of 33.4 annotations per image. The remaining 100 images (batch2) were sourced from a Kaggle dataset created by Buqi_, a student at the University of Chinese Academy of Sciences. These images were subsequently annotated manually by myself using the dataset trained on my prevouse images along with SAM.

Preprocessing and Data Augmentation: The original images (batch1) were auto-orientated and resized to 640x640 pixels. Several data augmentation techniques were applied, generating three outputs per training example. These augmentations include:

Rotation: Between -22° and +22°
Saturation: Between -85% and +85%
Brightness: Between -25% and +25%
Noise: Up to 5% of pixels
Mosaic: Applied
90° Rotate: Clockwise, Counter-Clockwise, Upside Down
Blur: Up to 1.5px

After applying these augmentations, the dataset expanded to 584 images with 16 null examples (mostly of the environment in which this is to be used), comprising 12,943 annotations with an average of 54.6 annotations per image. The average image size reduced to 0.41 MP, and the median image ratio became 640x640 square.

Annotation Distribution: The annotations were distributed across the four classes as follows:

Bolt: 4,197
Nut: 4,050
Washer: 2,511
Screw: 2,185

By using RoboFlow, we aim to train a robust object detection model to recognize and classify these hardware components accurately. This will enable efficient automation in tasks such as component identification, inventory management, and quality control in various industries.

the additional images from batch 2 have not been completed and added to the dataset at this time.