PSI - Issue 70

Anjireddy Mummadi et al. / Procedia Structural Integrity 70 (2025) 417–423

419

Faster R CNN framework by combining RPN and Fast R CNN, resulting in an integrated module that guarantees both speed and accuracy in detecting objects. 2.3. Mask R CNN By adding instance segmentation, Mask R-CNN expands on the capabilities of Faster R-CNN and allows for accurate object identification at the pixel level. This method, which was created by He et al. (2017), adds a specific segmentation branch that creates masks for distinct objects. To ensure precise segmentation, the technique uses a Fully Convolutional Network (FCN) to generate binary masks for every Region of Interest (RoI). Additionally, RoI Align, an improved pooling method intended to increase mask precision, is integrated into Mask R-CNN. This model's efficacy has led to its widespread use in high-accuracy applications like medical imaging, autonomous driving, and others. Based on existing literature , Faster R CNN and Mask R CNN have demonstrated superior performance in bolt loosening detection through object detection techniques. Therefore, this study focuses on evaluating these two methods to assess their effectiveness in detecting bolt loosening in steel structures . 3. Methodology

Image data set preparation

Data set (Roboflow)

Development of an RCNN model

Model training

RCNN model

Model validation

Faster RCNN

Segmenting the image and creating bounding boxes

Model comparison

Mask RCNN

Fig. 1. Flow chart depicting the methodology

4. Data set preparation

The methodology followed is presented in the Figure 1. The literature review is done to know present scenario and new techniques in the object detection. Quality of dataset is the key for the performance of deep learning. Supervised learning models require well annotated dataset. In this study, a dataset comprising 120 images was initially collected and then augmented to 600 images through the transformations like rotation, flipping, cropping, colour space alterations and noise injectio), covering two distinct bolt states: tight and loosened. To enhance dataset diversity, bolts were loosened by various rotations, and images were captured from different shooting angles and under varying lighting conditions.