Semantic segmentation, which predicts a semantic label for all pixels in the given image, plays a vital role in applications such as autonomous driving and robot navigation. However, most state-of-the-art models have low speed and require a large memory storage.
The proposed Context-aware Pruning (CAP) framework performs network pruning to remove the parameter redundancy of deep neural networks for semantic segmentation.
Different from existing pruning methods, the proposed framework exploits contextual information to guide the channel pruning. The original models can be significantly compressed leading to faster runtime and lower memory requirements, while preserving the original accuracy.
Technology Features, Specifications and Advantages
Given an existing deep neural network and training/testing dataset, the proposed framework automatically learns the contextual information and exploit this information to remove the parameter redundancies. The resulting model will be able to run faster and require less memory storage, while still preserving the accuracy of the original models.
The proposed framework can be used to compress various deep convolutional neural networks to run efficiently on edge devices, especially those targeted for computer vision tasks that rely heavily on contextual information, e.g., semantic segmentation and crowd counting.
The potential application areas include autonomous driving and robot navigation, which have strict requirements on speed, memory storage and power consumption.
The proposed framework enables the deployment of state-of-the-art deep neural networks on edge devices that are often battery powered, and have limited computational and memory resources, e.g., embedded systems, wearable and mobile devices.
The proposed framework enables existing complex deep neural networks to be deployed on low-cost embedded devices, while still achieving satisfactory performance.