Вы находитесь на странице: 1из 4

Pavement Distress Detection and Classification Using Image Classifier

What problem you were trying to solve?

Roads are an important infrastructure for the transportation of people and goods from
one place to another. To serve this purpose, it is required to design appropriate pavement to
prevent distress/defects/failures. Here, I am trying to identify the various distresses in the
pavement so that necessary maintenance strategy can be adopted to prevent it from failure.

What data did you collect?

I have collected images of pavement in three parts i.e. pavement without any distress
(the regular/normal ones), pavement with cracks on the surface and pavement with potholes
on the top layer. The websites GitHub and Kaggle were helpful in collecting the required data
for this study. The dataset contains 813 images for various classes of distresses, the example
of each class is represented below,

(a)

(b) (c)

Fig 1. Exemplary pavement images of different categories of


distresses from dataset (a) pavement without distress (b) pavement
with cracks on the surface (c) pavement with potholes
What classes did you use to classify your data and why they were chosen?

The various classes chosen are pavement without any distress (the regular/normal ones),
pavement with cracks on the surface and pavement with potholes on the top layer. These
classes were chosen because the pavement with various distresses requires different
maintenance strategies. For instance, pavement with cracks requires sealing and pavement
with potholes requires filling with the new bituminous mix. It is essential to identify the type
of distresses which leads to an increase in vehicle operating costs and accident rate due to
the movement of traffic.

How well did your classifier work? Were you happy with the results? What was the overall
percentage accuracy on test data?

In this study the classifier worked as desired; however, data cleaning is a bigger task
than inputting the data into a classifier and producing the output. I am happy with the results
obtained as the accuracy calculated using the confusion matrix was 95.6% and the precision
was 98.3%.

Accuracy: Overall, how often is the classifier correct?

(𝑇𝑃 + 𝑇𝑁) (235 + 4)


𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = = = 0.956
𝑇𝑜𝑡𝑎𝑙 250

where, TP – true positive and TP – true negative From the above equation it can be seen that
the accuracy obtained was 95.6% and hence the around 95.6% times the classifier worked
correctly.

Precision: When it predicts yes, how often is it correct?

𝑇𝑃 235
𝑃𝑟𝑒𝑐𝑖𝑠𝑜𝑛 = = = 0.983
𝑝𝑟𝑒𝑑_𝑦𝑒𝑠 239

From the above equation it is evident that the precision in predicting a yes when it is
correct is high (98.3%). Table 1 below provides more details on the same.
Table 1. Confusion Matrix
No of test samples - Predicted Predicted Total
250 No Yes
Actual TN= 4 FP=4 8
No
Actual FN= 7 TP =235 242
Yes
Total 11 239 250

Describe what images your model classifies well and which they classify badly?

Almost all the images collected from the website were clear which the model was able
to classify. However, the images clicked by zooming in and out needed some attention. Only
the images for the pavement with distresses were clicked by zooming in to capture the
distresses. A slight reduction or increase in the area that was zoomed for certain images were
the ones that the classifier was not able to classify. Examples of misclassified images for
different distresses are represented below,

Fig 2. Misclassified images for different classes are shown on the


right side and respective images from training dataset are shown on
the left side
Explain why you think it performed well or badly on the images you described in the last
part?

The images of the normal pavement were clicked along the length of the pavement.
The image captured having single lane or two lanes, during the ongoing construction, the
movement of pedestrians walking, the movements of vehicles, footpaths, and road markings.
In this study, images of two different types of pavement was captured which includes earthen
and paved pavements. The classifier failed to detect the earthen pavement and the road
under construction having the shortest length. The trained model had all the roads of greater
length.

In the second part, images were collected related to the pavement cracks. The
pavement cracks were zoomed to capture the distresses. The classifier failed because the
pavement cracks captured near the road markings, or near movement of pedestrians and
movement of vehicles was not able to detect as this was misleading the trained model with
the normal pictures.

The third part, deals with the images having potholes (depressions on the pavement).
The potholes were zoomed and captured. Here, the trained model was not able to classify the
images because it was trained with the depth of depression of potholes. The images with flat
depressions and larger area were the ones that were not classified correctly.

Were there problems with the classifier that you were able to solve? Describe your strategy
for solving the problem

As such there were no problems. The classifier could read only the pixels value, it would be
better if it could analyze the dimensions of the picture to classify, which requires optimization.
In this case convolution neural networks work better as they are complex and give high
accuracy with learnable weights and bias which enables us to classify images by
differentiating between each other.

Вам также может понравиться