Abstract:
"
Steel Frame Structure Defect Detection is one of the main stages in constructing a
building, where most of the times it has been done manually, which leads to no proper
inspection, on top of that there isn’t any technical device to monitor it, especially in Sri
Lanka. The aim of this dissertation is to detect six main defects in welded steel frame
structure by using image processing and deep learning algorithms, where the
application would aid individuals in construction site to identify defects in said steel
frame structures at an early stage of building in order to avoid casualties caused by the
defect. An android application incorporated with a classification model was proposed
and built. In this research, MobileNet has been used as the classifier algorithm, where
Transfer Learning has been implemented on the pretrained model on ImageNet. CNN
layers have been customized where GlobalAveragePooling2D layer has been
implemented with Rectified Linear Unit being the activation layer and being fed into
SoftMax layer. Furthermore, SGD optimizer with Categorical Cross Entropy Loss
functionality have been applied. An image preprocessing of data augmentation and
image transformation have been done. Viewpoint range is achieved at 10 – 3cm and
error free under device rotation circumstances. Robustness and processing performance
of the application have been achieved to an optimum level since it runs locally. The
mean accuracy level of device has been achieved for 91% for scratch, 78% for patches,
81% pitted surface, 78% for crazing, 73% for rolled in scale and 67% for inclusion
defects in welded steel frame structure which sums up with a model mean accuracy
being at 78%."
