Improving Business Processes with AI and Machine Learning
Weather patterns are changing. Every year, communities around the globe experience periods of some type of extremes: temperature, liquid and solid precipitation and wind. These extremes are particularly hard on buildings which get little protection from the surrounding areas.
Our Client performs inspection services in a market with a very active real estate market. The business is small but includes individuals, real estate agents and some work with insurance companies. The owner has always been safety conscious and looks for ways to improve the way the business performs its services.
Several discussions later, a Project Charter was submitted to the Client to use the bank of pictures they had taken at various sites to create a software tool that assesses the condition of roof shingles. A system would be developed that uses pictures to classify the condition of the shingles. The result would become part of the inspector’s assessment and report. To make this work, we would have to work through a series of problems to get a reliable working model: 1) Work with a limited dataset from the Client; 2) train, validate, test an image recognition model tailored to this application and 3) Perform all this without the use of large computing resources typically required for Big Data.
To do this, we used an existing ConvNet model containing hundreds of pictures of roof shingles with the expectation that it would be enough to get good accuracy. A ConvNet, Convolutional Neural Network (CNN), is a class of artificial neural networks, most commonly applied to analyzing visual imagery. After running the model with test data though, we achieved an accuracy of only 83% - Not accurate enough.
To overcome the obstacles, we researched a technique known as Transfer Learning to “teach” the existing ConvNet model wear patterns chosen from the Client’s dataset of a couple hundred pictures from past roof inspections. The resulting tailored model, based on the Client’s image dataset, would be created and then used to classify (none, low, moderate, severe) the wear of the roof shingles based on pictures taken at site.
The Transfer Learning improved the accuracy of the model from 83% to 91% ! Not bad considering that studies show that humans remember pictures a few days post-exposure with at least 90% accuracy. We consider that we have a useful model.
Below is a sketch depicting how the model works:
Investing more time and computational effort in training the original model to focus on even more relevant features could perhaps lead to further accuracy improvements up to 95% but with the model as is, there are already important benefits to moving forward:
- The model presents an objective evaluation of the wear increasing efficiencies to generate assessment reports;
- Using Robotic Process Automation (RPA), preliminary assessments can even be generated automatically with field pictures from a trusted source;
- The Client has an innovative toolset that improves its operations. More importantly, as more information is gathered, the toolset will become part of their augmented expert system;
- By using its available data effectively, this business has the power to transform itself to grow and match bigger players in its field.
Peter Darveau
