Automatic glass and concrete crack recognition system
An automatic system for inspection of the sky scrapper glass façade and concrete surfaces based on computer visions.
01
Histogram of Linearities
Linearity metrics are employed as the initial feature for distinguishing between images of cracked glass panels or concrete surfaces and those without cracks. For every crack edge segment, a value representing its linearity is computed. These linearity values are aggregated in a histogram and utilized for the purpose of classification
Published article:
Automatic Glass Crack Recognition for High Building Façade Inspection, click here
Published article:
A Generic Approach for Automatic Crack Recognition in
Buildings Glass Façade and Concrete Structures
02
Connected pixel configuration.
The next feature utilized for identifying cracks involves curvature like connected pixel configurations. Specific configurations of 3-connected pixels are counted in a histogram along the segments generated by an edge-detector. This histogram, representing the distribution of connected pixels throughout the entire image, is then employed to classify surfaces as either cracked or non-cracked
03
HOG based Method
In contrast to the previous approaches, no edge detector is used. Instead, it divides the input image into equal 3x3 blocks and computes a histogram for each of these blocks. Within each block, the partial derivatives in both the vertical and horizontal directions are computed for individual pixels. The magnitudes and orientations of these derivatives are then utilized to construct the histogram. The resulting 9 histograms from the 9 blocks are combined and provided as input to a classifier.
04
Partition based ULBP method
This technique computes the Uniform Local Binary Pattern (ULBP) code for each input image. Similar to HOG approach, the input image is divided into 9 uniform blocks. For each block, a histogram of ULBP values is generated. These 9 histograms are normalized and then combined by concatenation before being used as input to a classifier
05
CNN based Method
Given the success of Convolutional Neural Networks (CNNs) based on transfer learning in various machine learning tasks, a similar approach is adopted here for the recognition of cracks in glass and concrete. The model is constructed using the VGG16 and ResNet50 architectures, both of which have been pre-trained on the ImageNet dataset containing 1000 classes of natural images.
The transfer learning by feature extraction is used which uses the convolutional layers for extracting distinctive features without the need for retraining. These features are then directed to newly-constructed fully-connected layers, which are specifically designed for the purpose of classifying images with cracks. The fully connected layers are then trained on the cracked and non-cracked images