Chih-Hsien Hsia

A complexity reduction method for video synopsis system

This work proposes an efficient video retrieval technique for video synopsis. In order to reduce the search time, a new video synopsis search approach, low-complexity range tree algorithm, is proposed to improve the effectiveness of the search of the objects to match the requested conditions. With the time and space redundancy-reducing techniques of video synopsis, the objects searched by the users can be displayed in a short time. Therefore the objects and events can be found and displayed quickly without wasting time to watch those non-ROIs. For the test video sequences, the result shows an accuracy rate of 97% and a processing speed of 32 FPS (frame per second) of online phase, and the time complexity of object searching reduces from O(N) to O(logD-1N).

Spatial domain complexity reduction method for depth image based rendering using wavelet transform

Depth Image Based Rendering (DIBR) is an approach to generate a 3-D image by the original 2-D color image with the corresponding 2-D depth map. Although DIBR is a quite convenient technique of converting 2D to 3D images, there is a big problem in DIBR system that it cannot reach real-time processing due to the computing time. Therefore, this paper proposes a method based on discrete wavelet transform and adaptive edge-oriented smoothing process to improve the computing time of the system. The proposed method also preserves the original texture. As a results, this indicate that the proposed method not only preserves the vertical texture but also reduces at least 60% of the computing time of the DIBR system.

Finger-vein recognition based on parametric-oriented corrections

The two key factors in a biometric identification system are its high identification rate and convenience of device usage. In a finger-vein identification task, these two problems often occur since the captured device of finger-vein image should accommodate the high identification rate as well as the easy-to-use device design. The finger-vein is visually invisible inside the human skin. This work develops a new finger-vein capturing device using Near-Infrared (NIR) LED light and proposes an efficient technique for finger-vein identification. The vein image may contain noise and shadows due to device lighting conditions. Parametric-Oriented Histogram Equalization (POHE) is utilized to enhance image contrast and reduce the noise effect. This work also discusses normalized issues related to the angle correction of the finger edge and Region of Interest (ROI) for width normalization. In the experimental result, the proposed method yields a clear finger-vein pattern with a superior identification rate in the recognition task compared to the state-of-the-art methods.

A 3D Endoscopic Imaging System With Content-Adaptive Filtering and Hierarchical Similarity Analysis

In recent years, 3D endoscopic imaging has played an important role in the development of the medical industry, making it easy for medical doctors to judge the visual depth of ulcers or tumors and thereby decreasing the risk of unnecessary damages to other tissues. A traditional 2D endoscopic image can be converted into a 3D view through depth estimation and the use of a depth image-based rendering method. However, the application of depth estimation may be too time-consuming for use with current medical equipment. With the current equipment, the estimated depth map may also be drastically altered, causing the converted view to have unavoidable defects such as geometric distortion and hole effects. As a result of these dramatic changes, the depth map will have holes of different sizes and directions. Therefore, this paper presents the techniques of hierarchical similarity analysis and content-adaptive filtering as a means to transform the 2D image into a 3D view quickly and precisely. In order to reduce the computing time, this paper adopts hierarchical similarity analysis to separate images into different sizes in order to analyze the similarity and depth information estimation. After depth map estimation, a preprocessing procedure that applies respective Gaussian filters to different sizes and directions of holes is used to achieve the optimization of the depth map. As a result, this method not only reduces the system computation time, making it appropriate for use with current medical equipment, but also improves the 3D view quality for medical doctors.

Analyses of basketball player field goal shooting postures for player motion correction using kinect sensor

This work proposes an approach using the lowcost Kinect sensor as an assisting device that not only uses a contactless device to capture images information but cooperates with the application of developing field goal shooting posture detection algorithms for basketball players. The algorithms classify the field goal shooting postures of players into three stages as pre-shot, mid-shot, and post-shot routines each time for analyses. The Kinect first sets the posture recognizing conditions through the data provided by professional players and analyzes the postures affecting the highest and lowest field goal shooting percentage the most and least data from the statistical results.

An Assisted Forklift Pallet Detection with Adaptive Structure Feature Algorithm for Automated Storage and Retrieval Systems

This paper is about automatically guided vehicle (AGV) system in the automated-storage-and-retrieval-system (ASRS). In ASRS, it usually uses AGV system to transport materials, because it not only efficient but can cost down logistic cost. However, the major problem of the application about AGV is how to find the position of the pallets due to the difficulties to locating the pallet position on a complicated factory environment. In this work, Haar like-based Adaboost scheme with adaptive structure feature of pallets algorithm to detect pallets is presented, and by combining direction weighted overlapping (DWO) ratio, it can avoid those non-optimal candidates in object tracking. The experimental result shows this method can remove most of the non-stationary background and can increase the average pallet detection rate by 95%.

3D endoscopic imaging systems with DIBR

3D endoscopic imaging has played an important role in the development of the medical industry, making it easy for medical doctors to judge the visual depth of ulcers or tumors and thereby decreasing the risk of unnecessary damages to other tissues. A traditional 2D endoscopic image can be converted into a 3D view through depth estimation and the use of a Depth Image-Based Rendering (DIBR) method. However, the application of depth estimation may be too time-consuming for use with current medical equipment. In order to reduce the computing time, this work adopts hierarchical similarity analysis to separate images into different sizes in order to analyze the similarity and depth information estimation. As a result, this method not only reduces the system computation time, making it appropriate for use with current medical equipment, but also provide a good 3D view for medical doctors.

DIBR with content-adaptive filtering for 3D view

Depth Image Based Rendering (DIBR) is a well-known process for producing 3D views, which can synthesize a virtual view from images and their associated depth information. However, in an image owing to the abrupt change of the scene, the depth image may change rapidly to make the converting view have holes with different sizes and directions, and subsequently it affects qualities of the 3D view and may take more time for hole-filling. This paper presents a content-adaptive filtering technique to modify the depth map. A preprocessing that applies respective filter to different sizes and directions of holes is used to achieve the optimization of the depth map. As results, this method not only improves the 3D view quality and maintains the texture information, but also reduces the number of holes and the computing time on hole-filling.