Case Study: Using Deep Learning Techniques for Commercial Product Success

Submitted by BDTI on Thu, 04/13/2017 - 01:00

Here at BDTI, we’re working to apply deep learning techniques to a wide range of applications. Deep learning can be extremely effective—if there’s the right combination of data and processing power. The challenge to success lies in understanding how much data is sufficient and how to process it efficiently. This is where BDTI’s expertise in algorithms and architectures delivers value to our customers.

Recently, BDTI was engaged to create a convolutional neural network (CNN) to classify items into 25 categories for an embedded industrial application using image sensor input. BDTI started with a close examination of images provided by the customer and a careful analysis of the fundamental environmental challenges in the application. BDTI then created a custom CNN design with very low computational demand and memory footprint, which was well suited to simple and efficient implementation on a low-cost embedded CPU. The CNN achieved 100% accuracy on the customer’s validation dataset. However, because the training data did not adequately represent real-world conditions, BDTI also created custom data augmentation techniques and made suggestions for the collection of additional data to ensure the success of the application in the field as a commercial product.

The 2017 Embedded Vision Summit, taking place May 1-3 in Santa Clara, California, offers an excellent opportunity to learn more about how deep learning techniques are enabling more effective use of computer vision. Register today for this unique event. Readers of InsideDSP may take a 10% discount on registration with promotional code NLID0413.

Add new comment

Log in to post comments