As discussed in 03 Machine Learning and Neural Networks [ ↵ ], using an existing transferred model, already trained to recognise general features in an image, means only the final layer of the network needs to be re-trained on the specific task required.
In this demonstration the final layer is trained to recognise the difference between a bottle, can or coffee cup; initially using 50 images of examples of each object.
Here is a walkthrough video and example of how to train the model and identify the initial examples of each object, 02 How to Use the Project Controls to Identify Objects
On pressing the Recognise button, the system makes a prediction of the likelihood that the image it sees belongs to each category from its training data, and the most likely category is used as the identification.
In this demonstration you can also investigate how well the identification works, and re-train on other examples of the objects, learning how to improve the performance by including more training images.
Here is the walkthrough video for 04 Improving the Performance for New Objects,
Written guides based on these walkthrough videos can be downloaded here,
02 How to Use the Project Controls to Identify Objects_Guide [ ⇓ ]
04 Improving the Performance for New Objects_Guide [ ⇓ ]
If you delve into the Python code, you can even try using different transfer models! [ ↵ ]
Quick Shortcuts to project resources:
a glossary of the highlighted technical terms used can be downloaded here [ ⇓ ]
download the eBooklet [ ⇓ ], a digital copy of the printed project booklet
Share this post
