The authors of the NPU-BOLT: A Dataset for Bolt Object Detection in Natural Scene Images delve into the realm of engineering structures, focusing on bolt joints, which are commonplace and pivotal. These joints are susceptible to loosening or disengagement due to extreme service environments and load factors, posing a critical challenge for structural safety and longevity. The authors acknowledge the urgent need for real-time or timely detection of such issues in practical engineering applications.

Recent years have seen the emergence of various bolt loosening detection methods utilizing deep learning and machine learning techniques. However, most of these methods rely on laboratory-captured bolt images, characterized by controlled lighting, distance, and viewing angles. Real-world bolt images, on the other hand, often exhibit complexities such as blurred edges, oblique perspectives, partial occlusion, and variable colors, rendering models trained in controlled lab conditions less accurate or ineffective.

There are three sources of image-sets: natural environment photos, photos from the internet, and the CAD renders of a simulation.

For the images captured by the authors in natural environment, the following practical conditions, different light intensity, shadow shielding, large area corrosion, bolt shedding, cracks, rain and snow weather are considered. The real images from natural scenes are captured by a Cannon 200D hand-hold camera, an Honor X10 mobile phone and a DJI AIR 2plus drone. The images of real bolt connections in sunny, cloudy, night and indoor (building) environments were collected. A variety of bolted structures are captured to cover as many as possible bolted joint types, which include bolted flange joints, pipe connections and different bolted structures from bridge steel frame, vehicle hub, construction machinery, guardrails, iron cable fixings and anchor bolts.

The reason for using the downloaded pictures from the internet is that if the images in the dataset are all from the same photographers, the photographers’ habits may reduce the diversity of the dataset. The images on the internet are generally captured by different photographers and covers many engineering objects. To include the images from internet can improve the application scope and diversity of the dataset. However, generally the pixel quality of these pictures is slightly lower than the pictures captured by the authors. And, most of the image sizes are between 500x500 and 1500x1500.

The reason to add 3D CAD images is that they usually have high resolutions and clear bolt features without background information. Beside, the view angle can be easily adjusted in the CAD model. With adding a few number of 3D CAD simulation images in the dataset, it is expected to enhance the bolt object detection accuracy.

The dataset comprises 337 samples of bolt joint images primarily obtained from the natural environment. These images encompass a wide range of data sizes, spanning from 400x400 to 6000x4000 pixels, with a total of around 1275 annotated bolt targets. The dataset categorizes the bolt targets into four classes: blur bolt (loss of camera focus), bolt head (head of the bolt), bolt nut (nut of the bolt) and bolt side (side of the bolt).