events

executive

news

Gradual Saliency Detection in Video Sequences Using Bottom-up Attributes

Dr. Jila Hosseinkhani

Postdoctoral Researcher and Instructor, Carleton University

September 24, 2020 14:30 - 15:30

This is an online event. The details on how to join the event will be available once you register.

Register at EventBrite

abstract

The demand for video streaming is growing every day which implies a higher demand for new video transmitting and compression techniques to avoid data traffics over telecommunication networks. In this dissertation, we studied saliency detection in order to apply it to video streaming problem to be able to transmit different regions of video frames in a ranked manner based on their importance (i.e., saliency). Salient areas are the regions of interest that stand out relative to their surroundings and consequently attract more attention. To determine the salient areas within a scene, visual importance and distinctiveness of the regions must be measured.

The lack of a comprehensive and precise biologically-inspired study on the saliency of bottom-up stimuli prevents justifying the level of importance for different stimuli such as color, luminance, texture, and motion on the human visual system (HVS). To overcome this barrier, we investigated the bottom-up features using an eye-tracking procedure and human subjects in video sequences to provide a ranking saliency system stating the most dominant elements for each feature individually as well as in combination with other features. The experiment was performed under conditions in which we had no cognitive bias in order to speed up the video streaming procedure. Next, we introduced a gradual saliency detection framework for both still images and video sequences using color, texture, and motion features (based on our experimental estimations). In our algorithm, we proposed new feature maps for color and texture features, and we also improved the optical flow field estimation in our motion map.

Finally, different feature maps were combined and classified as different saliency levels using a Naive Bayesian Network. This work provides a benchmark to specify the gradual saliency for both static and dynamic (i.e., moving backgrounds) scenes. The main contribution of this work is the ability to assign a gradual saliency for the entirety of an image/video frame rather than simply extracting a salient object/area, which is widely performed in the state-of-the-art.

biography

Jila Hosseinkhani received her B.Sc. degree in electronics from AmirKabir University of Technology (Tehran Polytechnic), Tehran, Iran, and the M.Sc. degree in telecommunication engineering from the University of Tehran, Tehran, Iran. She completed her Ph.D. degree in electrical and computer engineering from Carleton University, Ottawa, Canada, in 2020. Her research interests include signal/image/video processing, pattern classification, artificial intelligence, and machine learning.

During her M.Sc., she developed an object detection and tracking algorithm with the aim of event detection and video summarization in soccer videos. The algorithm performed efficient ball detection and tracking using the Kalman Filter. She was awarded a Grant because of her work prototype from the Iranian TV and Broadcasting Center. She also won a DAAD scholarship as a visiting researcher at Rostock University, Rostock, Germany. After completion of her M.Sc., she worked as a Software Developer and a Researcher at Iranian Telecom Research Center (ITRC), Tehran, Iran, for 2 years.

During her Ph.D., she researched under a grant from NSERC on saliency detection. She developed advanced algorithms to output a gradual saliency map (i.e., heatmap) that can classify the different levels of importance for different regions within a video frame. Different parts of the saliency detection algorithm were established using segmentation, feature extraction, motion detection, optimization, and Bayesian network concepts. She is currently a postdoctoral research associate and an instructor at Carleton working on computer vision and machine learning areas.