Towards More Efficient and Accurate Deep Learning Models for Segmentation, Classification, and Tracking

点击翻译按钮获取中文摘要

Semantic segmentation and classification tasks have become an integral part of several vision intelligence (VI) based applications, like robot navigation, computer-aided diagnosis, video surveillance, and automatic production monitoring. Robustness and performance of these applications largely depend on the aforesaid low-level visual perceptions: the dense labeling task that assigns a label to each pixel in the image, and the global labeling task that assigns a single label to the whole image. Although enormous efforts have been undertaken with the advancement of deep learning (DL) to improve the models performing these tasks, unfortunately, the best models are compute and memory intensive. For instance, the RefineNet [1] architecture contains more than 85 million parameters, needing a minimum of 340 MB of memory in a 4-byte system. Hence, such models require expert labeled big data and high-performance computing (HPC) devices, like specialized graphic processing units (GPUs) or tensor processing units (TPUs). Additionally, most of the existing algorithms handle the segmentation and classification tasks in independent learning pipelines. These limitations hinder the practical adoption of the models in resource-limited environments (low-compute capacity platforms or a condition of data scarcity) and safety-critical applications with strict inference latency. Thus, the proposed research program is timely, and its long-term goal is to develop innovative and cost-effective DL models that can jointly segment and classify visual representations, including images and videos, for accurate real-time perception. Hence, the proposed models should maintain economic memory usage, high inference speed, and remarkable performance, especially on resource-limited platforms, to be deployed for practical purposes. Thus, it will be a research-intensive development of cutting-edge algorithms leveraging the strength of spectral-domain analysis, graph network theories, and DL. In line with the long-term goal, the research program is divided into three distinct themes: 1. Developing advanced DL models for more accurate image segmentation and classification, jointly. 2. Building innovative simultaneous video object segmentation and tracking. 3. Proposing efficient strategies for DL model optimization. These objectives have multifaceted challenges and require HQP with significant background knowledge, expertise, and coding skills to contribute to the program and improve upon the existing methods. Thus, a key focus of this program is training the next generation of researchers, and the funding requested will support 1 Ph.D., 4 MASc., and 5 undergraduate students over the next 5 years. The HQP will develop valuable competencies in the rapidly growing areas of image processing, computer vision, and deep learning that are in high demand by both the industry and academia. Thus, the HQP trained through this program will impact significantly Canadian research and development.