Qiuqi Ruan

One-shot learning gesture recognition from RGB-D data using bag of features

For one-shot learning gesture recognition, two important challenges are: how to extract distinctive features and how to learn a discriminative model from only one training sample per gesture class. For feature extraction, a new spatio-temporal feature representation called 3D enhanced motion scale-invariant feature transform (3D EMoSIFT) is proposed, which fuses RGB-D data. Compared with other features, the new feature set is invariant to scale and rotation, and has more compact and richer visual representations. For learning a discriminative model, all features extracted from training samples are clustered with the k-means algorithm to learn a visual codebook. Then, unlike the traditional bag of feature (BoF) models using vector quantization (VQ) to map each feature into a certain visual codeword, a sparse coding method named simulation orthogonal matching pursuit (SOMP) is applied and thus each feature can be represented by some linear combination of a small number of codewords. Compared with VQ, SOMP leads to a much lower reconstruction error and achieves better performance. The proposed approach has been evaluated on ChaLearn gesture database and the result has been ranked amongst the top best performing techniques on ChaLearn gesture challenge (round 2).

Letters: Palmprint recognition using Gabor-based local invariant features

Variations occurred on palmprint images degrade the performance of recognition. In this paper, we propose a novel approach to extract local invariant features using Gabor function, to handle the variations of rotation, translation and illumination, raised by the capturing device and the palm structure. The local invariant features can be obtained by dividing a Gabor filtered image into two-layered partitions and then calculating the differences of variance between each lower-layer sub-block and its resided upper-layer block (called local relative variance). The extracted features only reflect relations between local sub-blocks and its resided upper-layer block, so that the global disturbance occurred on palmprint images is counteracted. The effectiveness of the proposed method is demonstrated by the experimental results.

Letters: Palmprint recognition using Gabor feature-based (2D)2PCA

In this paper, we propose a novel approach of Gabor feature-based (2D)^2PCA (GB(2D)^2PCA) for palmprint recognition. Three main steps are involved in the proposed GB(2D)^2PCA: (i) Gabor features of different scales and orientations are extracted by the convolution of Gabor filter bank and the original gray images; (ii) (2D)^2PCA is then applied for dimensionality reduction of the feature space in both row and column directions; and (iii) Euclidean distance and the nearest neighbor classifier are finally used for classification. The method is not only robust to illumination and rotation, but also efficient in feature matching. Experimental results demonstrate the effectiveness of our proposed GB(2D)^2PCA in both accuracy and speed.

Letters: Facial expression recognition based on two-dimensional discriminant locality preserving projections

In this paper, a novel method called two-dimensional discriminant locality preserving projections (2D-DLPP) is proposed. By introducing between-class scatter constraint and label information into two-dimensional locality preserving projections (2D-LPP) algorithm, 2D-DLPP successfully finds the subspace which can best discriminate different pattern classes. So the subspace obtained by 2D-DLPP has more discriminant power than 2D-LPP, and is more suitable for recognition tasks. The proposed method was applied to facial expression recognition tasks on JAFFE and Cohn-Kanade database and compared with other three widely used two-dimensional methods: 2D-PCA, 2D-LDA and 2D-LPP. The high recognition rates show the effectiveness of the proposed algorithm.

Secure semi-blind watermarking based on iteration mapping and image features

This paper presents a secure semi-blind watermarking technique based on iteration mapping and image features. An image (gray or color) is divided into blocks of fixed size that are analyzed using fractal dimension to determine their properties. The feature blocks containing edges and textures are used to form a feature label. The watermark is the fusion of image feature label and a binary copyright symbol. Arnold iteration transform is employed for constructing the watermark to increase the security. The secure image adaptive watermark is then embedded in the feature blocks by modifying DCT middle-frequency coefficients. The detection and extraction procedure is a semi-blind, i.e., it does not need the original image. Only those who have the original watermark and the key can detect and extract the right watermark, which makes the approach have high security level. Experimental results demonstrate that this algorithm can achieve good perceptual invisibility, adaptability and security. It is also robust against spatial attacks including discarding, scribbling, luminance and contrast changing, and other attacks including low or high-pass filtering, adding noise and JPEG processing.

Shift and gray scale invariant features for palmprint identification using complex directional wavelet and local binary pattern

In this paper, a novel feature extraction framework is presented for palmprint identification, which provides a shiftable and gray scale invariant description of image achieving high identification accuracy at a low computational cost. The image is firstly decomposed by the shiftable complex directional filter bank (CDFB) transform which provides a two-dimensional (2-D) decomposition of energy shiftable and scalable multiresolution, arbitrarily directional resolution, low redundant ratio, and efficient implementation. Further, the subband coefficients of CDFB decomposition are operated by the uniform local binary pattern (LBP) which is gray scale invariant and contains information about the distribution of the local micro-patterns. The resulting LBP mappings are divided into many subblocks, over which the statistical histograms are achieved independently. Finally, a Fisher linear discriminant (FLD) classifier is learned in the statistical histogram feature space for palmprint identification. Experiments are executed over the HongKong PolyU palmprint database of 7752 images. To verify the high performance of our proposed feature descriptor, several other multiresolution and multidirectional transforms are also investigated including Gabor filter, dual-tree complex wavelet and Contourlet transforms. The experimental results demonstrate that CDFB yields the most promising performance balancing the identification accuracy, storage requirement and computational complexity for our proposed feature extraction framework.

Pinpoint authentication watermarking based on a chaotic system.

Watermarking technique is one of the active research fields in recent ten years, which can be used in copyright management, content authentication, and so on. For the authentication watermarking, tamper localization and detection accuracy are two important performances. However, most methods in literature cannot obtain precise localization. In addition, few researchers pay attention to the problem of detection accuracy. In this paper, a pinpoint authentication watermarking is proposed based on a chaotic system, which is sensitive to the initial value. The approach can not only exactly localize the malicious manipulations but reveal block substitutions when Holliman-Memon attack (VQ attack) occurs. An image is partitioned into non-overlapped regions according to the requirement on precision. In each region, a chaotic model is iteratively applied to produce the chaotic sequences based on the initial values, which are determined by combining the prominent luminance values of pixels, position information and an image key. Subsequently, an authentication watermark is constructed using the binary chaotic sequences and embedded in the embedding space. At the receiver, a detector extracts the watermark and localizes the tampered regions without access to the host image or the original watermark. The precision of spatial localization can attain to one pixel, which is valuable to the images observed at non-ordinary distance, such as medical images and military images. The detection accuracy rate is defined and analyzed to present the probability of a detector making right decisions. Experimental results demonstrate the effectiveness and advantages of our algorithm.

3D Facial expression recognition based on basic geometric features

This paper describes a 3D facial expression recognition approach based on distance and angle features, which can be got from the localized facial feature points. The probabilistic Neutral Network (PNN) architecture is used to classify the facial expressions based on BU-3DFE database. This paper adds the facial feature vectors with the information of slopes and the angles as the feature vectors got from the facial feature points, not only the distance information mentioned in the previous work. Thus it receives a better performance with an average recognition rate of 90.2%.

An illumination normalization model for face recognition under varied lighting conditions

In this paper, a novel illumination normalization model is proposed for the pre-processing of face recognition under varied lighting conditions. The novel model could compensate all the illumination effects in face samples, like the diffuse reflection, specular reflection, attached shadow and cast shadow. Firstly, it uses the TV_L^1 model to get the low-frequency part of face image, and adopts the self-quotient model to normalize the diffuse reflection and attached shadow. Then it generates the illumination invariant small-scale part of face sample. Secondly, TV_L^2 model is used to get the noiseless large-scale part of face sample. All kinds of illumination effects in the large-scale part are further removed by the region-based histogram equalization. Thirdly, two parts are fused to generate the illumination invariant face sample. The result of our model contains multi-scaled image information, and all illumination effects in face samples are compensated. Finally, high-order statistical relationships among variables of samples are extracted for classifier. Experimental results on some large scale face databases prove that the processed image by our model could largely improve the recognition performances of conventional methods under low-level lighting conditions.

3D SMoSIFT: three-dimensional sparse motion scale invariant feature transform for activity recognition from RGB-D videos

Abstract. Human activity recognition based on RGB-D data has received more attention in recent years. We propose a spatiotemporal feature named three-dimensional (3D) sparse motion scale-invariant feature transform (SIFT) from RGB-D data for activity recognition. First, we build pyramids as scale space for each RGB and depth frame, and then use Shi-Tomasi corner detector and sparse optical flow to quickly detect and track robust keypoints around the motion pattern in the scale space. Subsequently, local patches around keypoints, which are extracted from RGB-D data, are used to build 3D gradient and motion spaces. Then SIFT-like descriptors are calculated on both 3D spaces, respectively. The proposed feature is invariant to scale, transition, and partial occlusions. More importantly, the running time of the proposed feature is fast so that it is well-suited for real-time applications. We have evaluated the proposed feature under a bag of words model on three public RGB-D datasets: one-shot learning Chalearn Gesture Dataset, Cornell Activity Dataset-60, and MSR Daily Activity 3D dataset. Experimental results show that the proposed feature outperforms other spatiotemporal features and are comparative to other state-of-the-art approaches, even though there is only one training sample for each class.