The age of manual data manipulation has long passed and ANB Systems has joined many other companies in the race to embrace Machine Learning and AI as they dominate the near future. With image and optical character recognition as crucial pillars of our MLAI set-up, we have never been more equipped to break through the restraints of manual, non-automated processes.
Here are some ways we’re filling the gap:
Rather than generate a machine learning model from scratch, which takes more resources and is likely less accurate, we take a more refined approach by using an existing model and building layers on top of it. Open sourced models such as those created in coordination with Tensorflow (including MobileNet, ImageNet etc) contain data on image classification and can accurately classify a wide variety of everyday objects. However, at the lower levels, these models contain layers upon layers of neural networks whose characteristics can be generalized and thus molded into specific use cases. For example, ANB Systems is streamlining its data organization process by categorizing images of gauges into a screenshot class and camera-taken image class. In doing so, we utilize the ImageNet model which decreases the total amount of code and sample images necessary while increasing accuracy.
Sample images drive model creation. Without enough images for training, it is difficult to create models that are specific and accurate to fulfill our use cases. Here, image augmentation can be extremely beneficial. Image augmentation is the generation of sample images by slightly modifying existing images. This can be done via cropping, color-grading, mirroring, bleaching, and rotating images. This creates variation for models to be able to take in a wider variety of images.
Browser-Side Machine Learning
This article is written by Aneesh Gupta, Summer 2020 Intern from Duke University. To know more about DRM or other ANB Products, please write to firstname.lastname@example.org