Image Analytics with Machine Learning
Image analytics, deep learning and neural networks are terms thrown around all day on platforms like LinkedIn, but what can they really do for your business? Here at Altis, we have been working to provide an answer to that problem. Sure, the technology is really cool, but it would be even cooler with a project with some real ROI potential behind it.
Recently we have been working on an image analytics project which uses advanced neural networks to determine if the objects in the picture require cleaning or not. This is a potentially huge ROI as the client was spending thousands on cleaning their network without all of it needing to be cleaned!
We worked with their team to build a successful model that takes images of these objects and flows them into a neural network that takes the features of the image and analyses them for consistencies. The network we employed is called a ‘feed forward multilayer perceptron’ network, however it’s much easier to think of it using the system it was designed on: neurons in your brain! What it does is each ‘node’ in the neural network (NN) is used to recognise or discriminate a particular pixel of the image, these then all work together to from a consensus to classify the image much like the way we see an object and determine what it is from its characteristics.
This neural infrastructure model was built on TensorFlow, a highly advanced deep learning library developed by Google. This enabled us to create reliable and detailed neural infrastructures, as well as tailor and modify them as needed. As this was quite a specific use case for image analytics, it required the team to build the model from the ground up, circumventing any inbuilt training that most image analytic systems have.
Using tensorflow we developed the model, as shown above. This model has three layers, an input layer (the training data), hidden layer (the ‘nodes’ or ‘neurons’) and the output layer (which is the classification). The data is passed into the model by converting it to match the input layer (we used 1000 * 1000 pixels, quite large!) and then the hidden layer is created using the neural network discrimination parameters we developed.
When the model was built we trained it on the images provided by the client to create a working predictive analytics model. This model was able to predict successfully 74% of the cases where an object did or did not need cleaning. This can potentially be used to drastically reduce the amount of unnecessary cleaning done on the network, saving money and re-directing resources to other higher value (Customer satisfaction) work. The next steps planned will include the automated monitoring of the network from IoT cameras.
At Altis we aim to connect state of the art technologies to your real world business problems, using advanced techniques to solve problems with real benefit. Contact us and see what we can do for your business!