Applied Mechanics and Materials Vols. 571-572

Abstract: A deep Neural Network model was trained to classify the facial expression in unconstrained images, which comprises nine layers, including input layer, convolutional layer, pooling layer, fully connected layers and output layer. In order to optimize the model, rectified linear units for the nonlinear transformation, weights sharing for reducing the complexity, “mean” and “max” pooling for subsample, “dropout” for sparsity are applied in the forward processing. With large amounts of hard training faces, the model was trained via back propagation method with stochastic gradient descent. The results of shows the proposed model achieves excellent performance.

Abstract: A novel three-dimensional (3D) convex hull method is proposed, which is called dimensionality reduction convex hull method (DRCH).Through having 3d point set map to 2d plane, most initial 3D points in the convex hull are removed. Then, the remaining points are to generate 3D convex hull using any convex hull algorithm. The experiment demonstrates 3D DRCH is faster than general 3D convex hull algorithms. Its time complexity is O(r log r), where r is the number of points not in the hull. And DRCH can be generalized to higher-dimensional problems.

Abstract: This paper presents an enhanced multiple instance learning (MIL) tracker with a suitable representation of object appearance called distribution field descriptor (DF). To address transformations of object template (rotation, scaling), we firstly replace the smoothed histograms used in DF with smoothed bins according the theory of averaged shifted histograms. Secondly, due to the DF specificity and landscape smoothness, we adopt DF descriptor instead of traditional Haar-like one to represent the object appearance. By build object model using selected discriminative layers, our tracker is more robust while needing fewer features than the original tracker. The experimental results show higher performances of our tracker than those of five state-of-the-art ones on several challenging video sequences.

Abstract: Normal vector of 3D surface is important differential geometric property over localized neighborhood, and its abrupt change along the surface directly reflects the variation of geometric morphometric. Based on this observation, this paper presents a novel edge detection algorithm in 3D point clouds, which utilizes the change intensity and change direction of adjacent normal vectors and is composed of three steps. First, a two-dimensional grid is constructed according to the inherent data acquisition sequence so as to build up the topology of points. Second, by this topological structure preliminary edge points are retrieved, and the potential directions of edges passing through them are estimated according to the change of normal vectors between adjacent points. Finally, an edge growth strategy is designed to regain the missing edge points and connect them into complete edge lines. The results of experiment in a real scene demonstrate that the proposed algorithm can extract geometric edges from 3D point clouds robustly, and is able to reduce edge quality’s dependence on user defined parameters.

Abstract: In this paper we propose a method of vision-based hand detection. Since the pose space of articulated is infinite and it is not feasible to train a classifier for the recognition of all hand poses to detect hand areas. Therefore we assume that it is rather better to detect only one dominant hand pose. To do so after hand skin color segmentation, each interest region of hand area in the given image is validated that the shape of hand region is similar to the dominant pose. The proposed system has the process of two steps; hand candidate detection and dominant pose recognition using EOH feature and SVM classifier. The experimental result shows that the proposed method works very effectively with very low false negative rate 0.6%.

Abstract: In this paper, we propose a newly developed 2D-LDA method based on weighted covariance scatter for face recognition. Existing LDA uses the transform matrix that maximizes distances between classes. In LDA, we have to convert from an image to one-dimensional vector as training vector. In 2D-LDA, on the other hand, we can directly use two-dimensional image itself as training matrix, so that the classification performance can be enhanced about 20% comparing LDA, since the training matrix preserves the spatial information of two-dimensional image. However 2D-LDA uses same calculation schema for transformation matrix and therefore both LDA and 2D-LDA has the heteroscedastic problem which means that the class classification cannot obtain beneficial information of spatial distances of class clusters since LDA uses only data correlation-based covariance matrix of the training data without any reference to distances between classes. In this paper, we propose a new method to apply training matrix of 2D-LDA by using WPS-LDA idea that calculates the reciprocal of distance between classes and apply this weight to between class scatter matrix. To evaluate the performance of proposed algorithm, we use the ORL face database that includes 40 people and 10 images individually. The experimental result shows that the discriminating power of proposed 2D-LDA with weighted between class scatter has been improved up to 2% than original 2D-LDA. This method has good performance especially when the distance between two classes is very close and the dimension of projection axis is low.

Abstract: The visual defects of the polymer polaroid have a direct and serious influence on the quality of TFT-LCD panels. A variety of image detection systems have been proposed and widely used by the manufacturers of polaroid and panels in order to detect the visual defects at the earliest possible stage in the production process. Some slight visual defects, however, are barely visible in the images acquired by a camera when under a common illumination condition. In order to deal with this problem, we present a novel machine vision system in which a stripe light source is introduced to illuminate the polaroid sample, and these special defects therefore become more visible. At the base of the aforementioned image enhancement, a straightforward and fast image processing algorithm is designed and implemented. The Morphology Template Method is investigated and the shapes, the locations and the sizes of the visual defects are extracted successfully. The experimental results demonstrate this methodology’s validity to inspect the visual defects of transparent multilayer polymer films.

Abstract: Denoising is an important issue for laser active image. This paper attempted to process laser active image in the low-dimensional sub-space. We adopted the principal component analysis with local pixel grouping (LPG-PCA) denoising method proposed by Zhang [1], and compared it with the conventional denoising method for laser active image, such as wavelet filtering, wavelet soft threshold filtering and median filtering. Experimental results show that the image denoised by LPG-PCA has higher BIQI value than other images, most of the speckle noise can be reduced and the detail structure information is well preserved. The low-dimensional sub-space idea is a new direction for laser active image denoising.

Abstract: Pedestrian detection is one of the critical benchmarks for object detection in computer vision. In recent years, more effective detectors and features, such as Histograms of Oriented Gradients (HOG) have been proposed. The process of HOG features calculation is slow, and the features cannot satisfy represent the human body. Therefore, we adopt the multi-channel features, and propose a new improved method for accelerated integral image, the execution time of which is less than the original method. In addition, we apply novel multi-scales detection to detect new scenario, which is based on the low-altitude UAV. Under such scenario our algorithm can handle the changing in pedestrian posture and occlusion cues. The experimental results indicate that our algorithm is rapid and efficient under dynamic camera, comparing with other methods in INRIA dataset.

Abstract: In order to solve the defect detection problems of black line and white line of QR Code. According to the linear properties of defect, this paper puts forward a kind of defect detection algorithm based on Hough Transform and vertical projection. Through the experiment testing, the accuracy of algorithm detection reached 98.57%, the average test time is 38.28ms. This algorithm can be transplanted to other types of QR code and industrial on-line detection system.