PSI - Issue 70

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

420

4.1. Hardware and software configurations The data set was captured using iPhone 14. The training of deep learning networks requires ample computing power. The workstation used featured 16 GB GPU with 11 GB of memory and software like tensorflow1.4, pytorch etc.The deep learning network used in this study are Faster RCNN and Mask R CNN with ResNet architecture run on detectron2. 5. ResNet based Object Detection In this work ResNet based algorithm is developed (Faster RCNN and Mask RCNN). The training process involves in data augmentation and regularization of data which includes cropping, flipping rotation, translation, colour space alterations and noise injection. The optimizer used in the precent work is Adam and SDG. Early stopping was implemented to improve the validation process. Mean Average Precision (mAP) and intersection over union (IoU)is calculated. The architecture of Faster and Mask RCNN are given in Figure 2. Table 1. Training parameters Parameter Value Model Architecture ResNet (Faster and Mask RCNN) Number of classes 2 (Loose, Tight) Optimizer Adam, SoftMax Learning Rate 0.001 Batch Size 32 Weight deacy 0.0001 Number of epochs 30

Fig .2. Architecture of Faster R CNN and Mask R CNN: RPN

A) A deep learning CNN model was developed using faster R CNN for object detection and mask R CNN for instance segmentation. Both architectures share a similar structure with three main components: feature extraction, region proposal networks (RPNs), and a region-based CNN (R CNN). ResNet combined with a Feature Pyramid Network (FPN) was employed. In Faster R CNN, these proposals are resized and processed through fully connected layers for object classification and bounding box regression. B) Mask R CNN extends this architecture by adding a fully convolutional mask branch, which performs pixel wise segmentation for each detected object. This mask branch outputs a binary mask for every region proposal, enabling Mask R CNN to achieve both object detection and precise instance segmentation simultaneously.