Hu Xuelong

Content-based Image Retrieval Using Fuzzy Hamming Distance

This paper addresses an effective issue of content-based image retrieval (CBIR) by presenting fuzzy Hamming distance (FHD). Firstly, the theory of FHD is introduced, which includes degree of difference and cardinality of the fuzzy set, and then follows defuzzification of FHD which is used as a fuzzy similarity measure. In the experiment, the image retrieval results show this method can be employed to capture implicitly some semantic contents of the image in addition to other features such as color, texture and shape. Finally, an implemented prototype system demonstrates a promising retrieval performance for a test database, as compared with FUZZYCLUB system.

A New Method for Semi-Automatic Image Annotation

Semantic-based image retrieval bridges the gap between visual features and human understanding of image in the field of image retrieval. Image annotation is one important technology of image retrieval based on the semantic. This paper proposed one method to realize semi-automatic image annotation with the tool Support Vector Machine (SVM). The image collection was divided into two parts, one for manual annotation and the other for testing. After being classified by SVM, the output was changed into a probability, and K-NN algorithm was applied to get the keywords for unlabeled images. The experiments show that the approach is feasible.

Saliency aggregation via hard-voting evolution

Generic object level saliency detection is important for many vision tasks, such as object detection, compression, recognition, segmentation and so on‥ A variety of saliency detection methods have been proposed in recently, which often complement each other. In order to combine them, we propose a Hard-voting Evolution saliency aggregation algorithm in this paper. Specially, it is consist of three stages. First, we integrate each individual map to get a reference map by linear summation. Second, each of the individual maps is updated in Bayes framework based on the obtained reference map. Third, we use our proposed approach to get the aggregated saliency map. Finally, we do iterative operation to further improve performance. Experiments on three publicly available datasets demonstrate that the proposed algorithm performs favorably against the state-of-the-art methods.

Optimum Watermarking Algorithm Based-on Cost Function

This paper proposed a cost function in DWT domain, which combines the two most critical properties including the transparency and robustness. Based on the cost function, the optimum location can be found for watermarking. Through modeling to the wavelet coefficient by the GGD (generalized Gaussian density), we proposed the method of estimation of shape parameter, which corresponding to LOD (locally optimum detection) also can be used in watermarking detection. By utilizing the proposed approach, a good compromise between the transparency and robustness can be gotten in experiments.

Research on Yangzhou Goose breeding system based on RFID

The establishment of electronic file for modern poultry production and breeding represents the new trend in the industry. At the same time, every step of poultry production has significant effect on the safety of poultry products, and as a result, technical methods need to be taken to ensure safety management, identify who is responsible and protect the brand reputation. This paper uses Yangzhou Goose as an example in studying the establishment of a standardized goose breeding basic database, design of RFID-based smart foot ring that can collect various kinds of information, in an effort to develop a smart goose breeding system. This smart goose breeding system can track information during breeding and production, which can offer individual production information and guarantee the safety of goose products while traditional recording system merely records pedigree information.

Robust compressive tracking based on SURF

Recently, more and more researchers focus on compressive tracking due to its simple and efficient performance. However, performance of classical compressive tracking algorithms are very weak in illumination change, rotation and motion blur, which produce the drift problems easily in object tracking. Therefore, we present a kind of compressive tracking algorithm based on SURF. Firstly, we extract the features for the appearancemodel from foreground and background of the target as the positive and negative samples using SURF; then, we compress the high-dimensional feature space to low-dimensional space on the basis of compressive sensing theory. Secondly, we classify the positive and negative samples according to naive Bayes classifier with online update. The experiments show that the method we proposed performs effective and efficient and runs in real time with an average of 27 fps on challenging test sequences.

Exemplar-based image inpainting algorithm using adaptive sample and candidate patch system

In the existing exemplar based image inpainting algorithms, the most similar match patches are used to inpaint the destroyed region, and they are searched in the whole source region in a fixed size. However, sometimes it would decrease the connectivity of structure and clearness of texture while increases the time complexity of this algorithm. To solve these problems, firstly it proposed an adaptive sample algorithm based on patch sparsity, it calculates the patch sparsity and through it dividing the patches location into three types. And then the size of the sample patch would be adaptively changed according to the type. Secondly it proposed a candidate patch system to improve the patch matching rate. From the result, we can see that the proposed method can match more significant patches than the traditional method, and it can give a better texture inpainting effect, especially when processing the images with complex and regular textures, and the image with large destroyed region.

Parameter estimation of wavelet coefficient for image based on fractal lower order statistics

An alpha stable distribution and its fractional lower order statistics (FLOS) are introduced in this paper. Meanwhile the non-Gauss characteristics of some signals are considered in the practical applications. The method based on FLOS is the important method of signal analysis and processing under the signal and/or noise condition of non-Gauss alpha stable distribution. The medicine images are taken as the example. The FLOS in company with wavelet transform is introduced, and two parameters of wavelet coefficients for the images: alpha and gamma are estimated. It is explained that when in the practical image processing, we must consider the non-Gauss characteristics of signal to avoid performance degeneration with conventional signal processing methods. The proposed algorithm builds the foundation for the following image processing such as using Bayes wavelet to remove the noise.