Lew Fock Chong Lew Yan Voon

Calibration of a three-dimensional reconstruction system using a structured light source

We present a method for calibrating a range finder system composed of a camera and a structured light source. The system is used to reconstruct the three-dimensional (3-D) surface of an object. This is achieved by projecting a pattern, represented by a set of regularly spaced spots, on the surface of the object using the structured light source. An image of the illuminated object is next taken and by analyzing the distortion of the projected pattern, the 3-D surface of the object can be reconstructed. This reconstruction operation can be envisaged only if the system is calibrated. Instead of using a classical calibration method, which is based on the determination of the matrices that characterize the intrinsic and extrinsic parameters of the system, we propose a fast and easy to set up methodology, consisting of taking a sequence of images of a plane in translation on which a set of regularly spaced spots is projected using the structured light projection system. Next, a relation- ship between the position of the plane and the coordinates of the spots in the image is established. Using this relationship, we are able to deter- mine the 3-D coordinates of a set of points on the objects surface know- ing the 2-D coordinates of the spots in the image of the object taken by the range finder system. Finally, from the 3-D coordinates of the set of points, the 3-D surface of the object is reconstructed.

Single tree species classification from Terrestrial Laser Scanning data for forest inventory

Due to the increasing use of Terrestrial Laser Scanning (TLS) systems in the forestry domain for forest inventory, the development of software tools for the automatic measurement of forest inventory attributes from TLS data has become a major research field. Numerous research work on the measurement of attributes such as the localization of the trees, the Diameter at Breast Height (DBH), the height of the trees, and the volume of wood has been reported in the literature. However, to the best of our knowledge the problem of tree species recognition from TLS data has received very little attention from the scientific community. Most of the research work uses Airborne Laser Scanning (ALS) data and measures tree species attributes on large scales. In this paper we propose a method for individual tree species classification of five different species based on the analysis of the 3D geometric texture of the bark. The texture features are computed using a combination of the Complex Wavelet Transforms (CWT) and the Contourlet Transform (CT), and classification is done using the Random Forest (RF) classifier. The method has been tested using a dataset composed of 230 samples. The results obtained are very encouraging and promising.

3D characterization of hot metallic shells during industrial forging

During industrial forging of hot metallic shells, it is necessary to regularly measure the dimensions of the parts, especially the inner and outer diameters and the thickness of the walls. A forging sequence lasts 2 h or more during which the diameter of the shell is regularly measured in order to decide when to stop the forging process. For better working conditions, for the safety of the blacksmiths, and for a faster and more accurate measurement, we have developed a novel system based on two commercially available time of flight laser scanners for the measurement of the diameters of hot cylindrical metallic shells during the forging process. The advantages of using laser scanners are that they can be placed very far from the hot shell, more than 15 m, while at the same time giving an accurate point cloud from which three-dimensional views of the shell can be reconstructed and diameter measurements done. Moreover, more accurate measurement is achieved in less time with the laser system than with the conventional method using a large ruler. The system has been successfully used to measure the diameters of hot cylindrical metallic shells.

Detecting parabolas in ultrasound B-scan images with Genetic-Based Inverse Voting Hough Transform

In this paper we propose a Genetic-Based Inverse Voting Hough Transform (GBIVHT) method for detecting parabolic shapes iii B-scan images obtained by the ultrasonic Time Of Flight Diffraction inspection technique. These parabolic shapes are characteristics of the presence of crack defects in the engineering structure under inspection. In our method, the local peak detection problem in the parameters space of conventional Hough Transform is converted into a parameter optimization problem that operates directly on the ultrasound B-scan image. The optimization is done using the well-known Genetic Algorithms. Our main goals are an accurate detection of the parabolas while circumventing the computational complexity and huge storage problem tied to conventional Hough Transform.

Hybrid segmentation of depth images using a watershed and region merging based method for tree species recognition

Tree species recognition from Terrestrial Light Detection and Ranging (T-LiDAR) scanner data is essential for estimating forest inventory attributes in a mixed planting. In this paper, we propose a new method for individual tree species recognition based on the analysis of the 3D geometric texture of tree barks. Our method transforms the 3D point cloud of a 30 cm segment of the tree trunk into a depth image on which a hybrid segmentation method using watershed and region merging techniques is applied in order to reveal bark shape characteristics. Finally, shape and intensity features are calculated on the segmented depth image and used to classify five different tree species using a Random Forest (RF) classifier. Our method has been tested using two datasets acquired in two different French forests with different terrain characteristics. The accuracy and precision rates obtained for both datasets are over 89%.

3D reconstruction of hot metallic surfaces for industrial part characterization

During industrial forging of big hot metallic shells, it is necessary to regularly measure the dimensions of the parts, especially the inner and outer diameters and the thickness of the walls, in order to decide when to stop the forging process. The inner and outer diameters of the shells range from 4 to 6 meters and to measure them a large ruler is placed horizontally at the end of the shell. Two blacksmiths standing on each side of the ruler at about ten meters from it visually reads the graduations on the ruler in order to determine the inner and outer diameters from which the thickness of the wall is determined. This operation is carried out several times during a forging process and it is very risky for the blacksmiths due to the high temperature of the shell when the measurement is done. Also, it is error prone and the result is rather inaccurate. In order to improve the working conditions, for the safety of the blacksmiths, and for a faster and more accurate measurement, a system based on two commercially available Time Of Flight (TOF) laser scanners for the measurement of cylindrical shell diameters during the forging process has been developed. The advantages of using laser scanners are that they can be placed very far from the hot shell, more than 15 meters, while at the same time giving an accurate point cloud from which 3D views of the shell can be reconstructed and diameter measurements done. Moreover, better dimensional measurement accuracy is achieved in less time with the laser system than with the conventional method using a large ruler. The system has been successfully used to measure the diameter of cold and hot cylindrical metallic shells.

Region-based segmentation on depth images from a 3D reference surface for tree species recognition

The aim of the work presented in this paper is to develop a method for the automatic identification of tree species using Terrestrial Light Detection and Ranging (T-LiDAR) data. The approach that we propose analyses depth images built from 3D point clouds corresponding to a 30 cm segment of the tree trunk in order to extract characteristic shape features used for classifying the different tree species using the Random Forest classifier. We will present the method used to transform the 3D point cloud to a depth image and the region based segmentation method used to segment the depth images before shape features are computed on the segmented images. Our approach has been evaluated using two datasets acquired in two different French forests with different terrain characteristics. The results obtained are very encouraging and promising.

Crack defect detection and localization using genetic-based inverse voting Hough transform

We propose a genetic-based inverse voting Hough transform (GBIVHT) method to detect buried crack defects in engineering structures. The method is applied to B-scan images obtained according to the ultrasonic time of flight diffraction technique. In these image representations of the ultrasound data, crack defects are characterized by multiple arcs of diffraction that can be approximated by a parabolic model. Thus, the crack defect detection problem in non-destructive inspection of engineering structures is transformed into a parabola detection and localization on B-scan images. In the proposed GBIVHT method, the local peak detection problem of conventional HT is converted into a parameter optimization problem that operates directly on the B-scan images. The optimization task is done using the well-known genetic algorithms. Our main goals are an accurate detection of the parabolas while circumventing the computational complexity and huge storage problem tied to conventional HT.

Silicon retina for real-time pattern recognition

We present in this paper a programmable silicon retina designed for real-time pattern recognition. Its working principle is based on the comparison between an image projected on the retina by some opt9ical means and a reference binary image or mask memorized in the circuit. The result of the comparison is two signals corresponding to the sum of the currents produced by the pixels pertaining to the black and white zones of the reference binary image, this image when projected on the retina will produce a maximum white pixel current and a minimum black pixel current if it coincides perfectly with the reference binary image. If the projected image is shifted with respect to the reference binary image or if it is different then the black and white pixel currents will be different also. By measuring these two currents and by comparing them to expected values, a shift of the pattern or a difference between the observed and programmed pattern can be detected. Extensive computer simulations have been done in order to validate the working principle of the retina. Moreover, in order to verify the feasibility of the circuit in CMOS technology, we have fabricated a prototype non-programmable circuit in 1.2 micrometers standard CMOS technology. The measurements done on this circuit are quite encouraging and have been found to correspond to our expectations. Finally, the architecture of the programmable silicon retina, designed in a more recent 0.6 micrometers CMOS technology, is presented. This circuit is currently being fabricated.

A CMOS retina for Zernike moments estimation

Indeed, if we consider the real and imaginary parts of the Zernick polynomial of order p and repetition q as two images, then we can notice that there is a close relationship between the correlation value of two images and the expression of the real and imaginary parts of the Zernick moments of an image. Thus, the value of the Zernick moment of an image can be obtained by computing the correlation value between the image under analysis and two other images, one for the real part and another one for the imaginary part. The latter two images that depend on the order p and repetition q of the Zernick moment to compute are gray level images that need to be memorized in the retina. In order to reduce hardware implementation cost they are transformed into binary images or masks using a dithering algorithm. In this way only a 2-bit memory device is required per pixel to memorize the two masks (on bit per mask). Using the binary masks instead of the gray level images only gives an approximate value of the Zernick moments. However, we will show that the approximated values are still a good representation of the analyzed image (and thus can be used in a pattern recognition application). To do so, the exact and approximate values of the Zernick moments for values of p and q ranging from 0 to 30 have been computed and the images reconstructed from these values compared to the original one. The relative errors between the respective reconstructed images (exact and approximated Zernick moments) and the original image have been plotted against the orders of the Zernick moments used in the reconstruction. We have noticed that the evolutions of the error curves are quite similar.