Webinar on Deep-Learning based Image Processing

UTAR organised a webinar titled "Deep-Learning based Image Processing" on 29 October 2020 via Zoom. The invited speaker for the webinar was UTAR International Collaborative Partner (ICP) Prof Hang Hsueh-Ming.

Prof Hang during the webinar

In his talk, Prof Hang first explained how the neural network mimics the works of the human brain. He then briefly introduced the human visual system and neutral cell connection, explaining how a human processes information. He then introduced the feed-forward neural network concept and explained how convolutions work in the convolutional neural network (CNN). He presented a well-known paper published by Zeiler and Fergus to illustrate how neural networks work. He mentioned that the goal of a neural network is to do classification, which involves recognising an image from low-level features to a trainable classifier.

Prof Hang then talked about semantic segmentation and he said it is popular for different applications such as self-driving. He mentioned that the purpose of semantic image segmentation was to classify every pixel of an image with a corresponding class of what is being represented. He also further explained the encoder-decoder structure. The structure, composed of an encoder and decoder, aimed to pass the spatial information in the encoder to the decoder.

Besides, he presented one of his recent research papers “Efficient dense modules of asymmetric convolution for real-time semantic segmentation”. The paper, which won the Best Paper Award in ACM-MM Asia conference, focused on designing a well-balanced network architecture for semantic segmentation. Prof Hang noted that relatively accurate solutions are now possible for its use in self-driving applications due to recent advancements, and the target of the research is to design a high-efficient system with high accuracy, real-time inference time and low model size.

Prof Hang then continued his presentation with another topic focusing on deep-learning based image edge detection. According to him, edge detection is a very popular and basic image processing technique, and deep neural network (DNN) based edge detection often has high accuracy performance. He shared another recent paper which focused on designing an effective and lightweight deep learning framework to detect edges. In this paper, a set of DNN based edge detectors, which are inspired by the traditional methods, was proposed to contribute a much lower model size at about the same edge detection accuracy.

While explaining the deep-learning based image compression, he shared another paper which focused on the hybrid layered image compressor with deep-learning technique. In his final remarks, he highlighted the efficiency of DNN in image processing, semantic segmentation and edge detection, as well as the potential of learned image compression to outperform traditional coding.

Prof Hang explaining the connections between the nodes

Prof Hang providing an example of semantic segmentation

Prof Hang comparing complexity with accuracy performance



Wholly owned by UTAR Education Foundation Co. No. 578227-M        LEGAL STATEMENT   TERM OF USAGE   PRIVACY NOTICE