Chuan-Min Zhai

Recognition of plant leaf image based on fractal dimension features

Abstract This work proposed a new method of describing the characteristics of plant leaves based on the outline fractal dimension and venation fractal dimension. First, using multiple threshold edge detection method to separate leaf edge and vein, and get multiple veins. Then, the two-dimensional fractal dimension of the leaf edge image and multiple vein images will be calculated, and a new ring projection wavelet fractal feature for leaf shape is also adopted. Finally, these two kinds’ fractal dimension features are used to plant leaves classification and recognition. The experimental results demonstrated the effectiveness of the fractal dimension feature method.

Face aging simulation and recognition based on NMF algorithm with sparseness constraints

Abstract In recent years, face recognition has been widely applied in managing and criminal fields. Apart from lighting, gesture and expression, variations in shape and texture of human faces due to aging factor would also affect the performance of face recognition systems extremely. A facial aging simulation method based on sparse-constrained method is first proposed and then applied in the age-across face recognition. Experiments show that age span indeed has a great effect on face recognition, but the recognition ratio is apparently improved after adding additional virtual samples by aging simulation.

Recognition of leaf image set based on manifold–manifold distance

Abstract Recognizing plant leaves has been a difficult and important work. In this paper, we formulate the problems by classifying leaf image sets rather than single-shot image, each of set contains leaf images pertaining to the same class. We extract leaf image feature and compute the distance between two manifolds modeled by leaf images. Specifically, we apply a clustering procedure in order to express a manifold by a collection of local linear models. Then the distance is measured between local models which come from different manifolds that constructed above. Finally, the problem is transformed to integrate the distance between pairs of subspace. Experiment based on the leaves (ICL) from intelligent computing laboratory of Chinese academy of sciences, which shows that the method has a great performance.

Deep Learning with PCANet for Human Age Estimation

Human age, as an important personal feature, has attracted great attention. Age estimation has also been considered as complex problem, how to get distinct age trait is important. In this paper, we investigate deep learning techniques for age estimation based on the PCANet, name DLPCANet. A new framework for age feature extraction based on the DLPCANet model. Different from the traditional deep learning network, we use PCA (Principal Component Analysis, PCA) algorithmic to get the filter kernels of convolutional layer instead of SGD (Stochastic Gradient Descent, SGD). Therefore, the model parameters are significantly reduced and training time is shorter. Once final feature has been fetched, we K-SVR (kernel function Support Vector Regression, K-SVR) for age estimation. The experiments are conducted in two public face aging database FG-NET and MORPH, experiments show the comparative performance in age estimation tasks against state-of-the-art approaches. In addition, the proposed method reported 4.66 and 4.72 for MAE (Mean Absolute Error, MAE) for point age estimation using FG-NET and MORPH, respectively.

Recognizing complex events in real movies by combining audio and video features

Abstract This paper proposes a novel approach in taking audio feature into account for better event recognition performance in recognizing complex events in real movies, since audio can provide strong evidence to certain events. In our method, local-space time feature and audio feature are firstly extracted from the video sequences and then an individual video sequence is represented as a SOFM density map; finally we integrate such density map with SVM for recognition events. To evaluate effectiveness of this method, this paper uses the public Hollywood dataset, in this dataset the shot sequences have been collected from 32 different Hollywood movies and it includes 8 event classes. The presented result justifies the proposed method explicitly, improve the average accuracy and average precision compared to other relative approaches.

Shape and Color Based Segmentation Using Level Set Framework

We propose a level set based variational approach that incorporates shape and color prior into Local Chan-Vese model for segmentation problem. Object detection and segmentation can be facilitated by the availability of a reference object. In our model, besides the level set function for segmentation, we introduce another labelling level set function to indicate the regions on which the prior shape and color should be compared. The active contour is able to find boundaries that are similar in shape and color to the prior, even when the entire boundary is not visible in the image. The experimental results demonstrate that the proposed model can efficiently segment the objects.

Integration of global and local feature for age estimation of facial images

Automatic age estimation from facial images is emerging as an important research area in recent years due to its promising potential for some computer vision applications. In this paper we propose a novel approach combine the global and local facial features in parallel manner to implement the age estimation. Then after extracting global and local features, these features are integrated for fine classification. In the proposed method, global and local features are extracted by Discrete Fourier Transform (DFT) and Principal Component Analysis (PCA) respectively. We have conducted experiments on a large scale age databases (FGNET). The experimental results are very promising in showing that it is an effective method

Event recognition based on a local space-time interest points and self-organization feature map method

This paper proposes a novel method based on local space-time interest points and self-organization feature map to recognize and retrieval complex events in real movies. In this method, an individual video sequence is represented as a SOFM density map then we integrate such density map with SVM for recognition events. Local space-time features are introduced to capture the local events in video and can be adapted to size and velocity of the pattern of the event. To evaluate effectiveness of this method, this paper uses the public Hollywood dataset, in this dataset the shot sequences has collected from 32 different Hollywood movies and it includes 8 event classes. The presented result justify the proposed method explicitly improve the average accuracy and average precision compared to other relative approaches.

Action recognition via an improved local descriptor for spatio-temporal features

This paper presents and investigates an improved local descriptor for spatio-temporal features on action recognition. Follow the idea of local spatio-temporal interest points on human action recognition, we develop a memory-efficient algorithm based on integral videos. The contribution of our job is we use the SURF descriptors on cuboids to speed up the computation especially for the integral video and improve the recognition rate. We present recognition results on a variety of dataset such as YouTobe and KTH, compared to previous work, the results showed that our algorithm is more efficient and accurate compared with the previous work.

Reverse Training for Leaf Image Set Classification

This paper presents a new approach for leaf image set classification, where each training and testing set contains many image instances of a leaf. This approach efficiently extends binary classifiers for the task of multi-class image set classification. First, the training set is divided into two part using clustering algorithms: one will train a classifier with the images of the query set; the rest of the training set will evaluate the trained classifier and then predict the class of the query image set. The PHOG feature and Gist feature of leaf image set are merged into the whole feature of leaf image sets. Extensive experiments and comparisons with existing methods show that the proposed approach achieves state of the art performance for leaf image set recognition.