Convolution Neural Networks (CNN) are rapidly deployed in ADAS and Autonomous driving for object detection, recognition, and semantic segmentation. The prior art of supporting CNN (HW IP or multi-core SW) doesn't address efficient implementation for the first layer, YUV color space,
and output stride support. The given paper proposes a new pre-processing technique to enhance CNN based HW IP or multi-core SW solution. The pre-processor enables new features namely (1) Higher parallelism for the first layer with boosting of first layer (2) Efficient YUV color space (3) Efficient
output stride support. The pre-processor uses novel phase-split method to enable supporting above features. The proposed solution splits input to multiple phases based on spatial location e.g. 2 phases for YUV 4:2:0 format, 4 phases for output strides 2 etc. The proposed solution is a unified
solution that enables utilization (>90%) for the first layer and reduction of bandwidth of 2-4x for output stride of 2. For YUV color space, this reduces the computation by factor 2 along saving of ∼0.1 mm2 of silicon area with negligible loss in accuracy.
No References for this article.
No Supplementary Data.
No Article Media
CNN (CONVOLUTION NEURAL NETWORK);
Document Type: Research Article
Publication date: January 29, 2017
More about this publication?
For more than 30 years, the Electronic Imaging Symposium has been serving those in the broad community - from academia and industry - who work on imaging science and digital technologies. The breadth of the Symposium covers the entire imaging science ecosystem, from capture (sensors, camera) through image processing (image quality, color and appearance) to how we and our surrogate machines see and interpret images. Applications covered include augmented reality, autonomous vehicles, machine vision, data analysis, digital and mobile photography, security, virtual reality, and human vision. IS&T began sole sponsorship of the meeting in 2016. All papers presented at EIs 20+ conferences are open access.
Please note: For purposes of its Digital Library content, IS&T defines Open Access as papers that will be downloadable in their entirety for free in perpetuity. Copyright restrictions on papers vary; see individual paper for details.