Ke Lv

A gradient descent boosting spectrum modeling method based on back interval partial least squares

When the technique of boosting regression is applied to near-infrared spectroscopy, the full spectrum of samples are generally used to perform partial least squares (PLS) modeling. However, there is a large amount of redundant information and noise contained in the full spectrum. This not only increases the complexity of the model, but also reduces its predictive performance. In addition, the boosting method is sensitive to data noise. When the data are mixed with too much noise, the generalization performance of boosting will decrease, and the prediction error and the variance of PLS will be relatively large. To solve these problems, a gradient descent boosting ensemble method combined with backward interval PLS (GD-Boosting-BiPLS) is proposed in this paper. BiPLS is used to select the effective variables for the boosting base model, and each base model is trained sequentially by resampling. The spectral segmentation parameter of BiPLS and the iteration parameter of boosting are fused, and the weight of each base model is distributed by the gradient descent strategy. This leads to a new ensemble model (forward additive model) in the direction of reduced residuals. The final model is the ensemble model that obtains the minimum root mean square error of prediction (RMSEP). The proposed method is applied to the quantitative prediction of ethanol concentrations. Over iterations 1-50, the average correlation coefficients of the calibration and validation sets are 0.9628 and 0.9388, and the average RMSE of cross-validation and RMSEP are 0.0732 and 0.0675, respectively. The overall performance of the proposed GD-Boosting-BiPLS method is compared with those of various ensemble strategies and 4 kinds of state-of-the-art spectral modeling methods. The experimental results reveal that the proposed method has the best generalization performance and stability. HighlightsAn ensemble model of gradient descent boosting and BiPLS is proposed.With BiPLS as the base model method can reduce the sensitivity of boosting to noise.The gradient descent boosting strategy can improve the performance of base models.The iteration parameter and the segmentation parameter are fused to simplify the model.The final ensemble model can remain stable at different initial number of iterations.

Large visual words for large scale image classification

Recently, using large visual vocabulary or codebooks to quantize and partition the set of local feature descriptors into large set of disjoint subsets termed visual words (or large visual words) has become an important research topic in solving many computer vision problems including near duplicate image retrieval, object retrieval, etc. Generally, large visual words means a heavy burden on the cost of time and memory space for both the construction of large vocabulary and the searching process, especially for large scale applications. In this paper, we present an efficient generation approach of large visual words with a very compact vocabulary, namely two dictionaries learned with sparse non-negative matrix factorization (NMF). After piecewise sparse decomposition of features with two learned dictionaries, we map a pair of indices of the dictionarys bases corresponding to the maximum elements of the two sparse codes to a large set of visual words upon the assumption that data with similar properties will share the same base with the largest sparse coefficient. With the help of an inverted file structure built through the large visual words, K-nearest neighbors (KNN) can be efficiently retrieved. Therefore, we can classify images very efficiently with the incorporation of our fast KNN search based on large visual words into SVM-KNN method. Experiments on the public Oxford dataset, and ACM Multimedia 2013 Yahoo! image classification challenge dataset show that our approach is both effective and efficient.

Foreground Detection Utilizing Structured Sparse Model via l1,2 Mixed Norms

Foreground object detection is a crucial technique of intelligent surveillance systems, and it is still a challenging problem in complex scenes with illumination variations and dynamic backgrounds. Intuitively, the foreground object pixels are often not sparsely distributed but tend to be clustered. Motivated by this hypothesis, we present a new structured sparse model to extract foreground objects, which introduces the spatial neighborhood information into a unified optimization framework by l1,2 mixed norms. Simultaneously, we also give the solving method of the proposed model in details. Moreover, we apply the model to the sparse signal recovery and background subtraction in videos. In the experiments, better performance is obtained over previous methods. The experimental results validate the hypothesis and the effectiveness of the proposed method.

Hamming embedding with fragile bits for image search

Recently, several binary descriptors are proposed, which represent interest points in image using binary codes. In these binary feature schemes, two descriptors are considered as a match, if the Hamming distance between them is below a threshold. Applying Hamming distance to measure the similarity between binary descriptors can extremely promote the computational efficiency. However, our experimental results presents that there exists a large number of bits in the binary feature vector cannot maintain the robustness while image conditions change. Rather than ignore the impacts of those unstable bits, we take into account the difference of robustness among the feature bits and propose a novel similarity measurement, which called the Fragile Bit Ratio (FBR). FBR is used in binary feature matching to measure how two features differ. High FBRs are associated with genuine matches between two binary features and low FBRs are associated with impostor ones. Based on this metric, we propose a new binary feature matching scheme to fuse the Hamming distance and Fragile Bit Ratio. In our approach, we match the descriptors using the Hamming distance threshold roughly, and then filtered by the Fragile Bits Ratio to refine the candidate set. In experiments, using Fragile Bits Radio can effectively remove the false matches and highly improve the accuracy of image search. Furthermore, our method can easily be integrated into the other well-established binary features schemes.

A fingerprint key binding algorithm based on vector quantization and error correction

In recent years, researches on seamless combination cryptosystem with biometric technologies, e.g. fingerprint recognition, are conducted by many researchers. In this paper, we propose a binding algorithm of fingerprint template and cryptographic key to protect and access the key by fingerprint verification. In order to avoid the intrinsic fuzziness of variant fingerprints, vector quantization and error correction technique are introduced to transform fingerprint template and then bind with key, after a process of fingerprint registration and extracting global ridge pattern of fingerprint. The key itself is secure because only hash value is stored and it is released only when fingerprint verification succeeds. Experimental results demonstrate the effectiveness of our ideas.

An improved cross-matching algorithm for fingerprint images from multi-type sensors

This paper proposes a novel cross-matching algorithm for fingerprint images from multi-type fingerprint sensors. Our method can handle the difference of fingerprint images which results by the different characteristics of fingerprint sensors. By using core detection based fingerprint registration and a two-level transformation - image space normalization and feature space normalization, all feature points of fingerprint images are mapped into one feature space. Then the feature points extracted in different sensor images are matched to calculate similarity in the same feature space. Experimental results show that better accuracy can be achieved on crossing matching image dataset after normalization. Our method presents the good potential on image datasets of optical sensors, thermal slice sensors and capacity sensors.

Unsupervised color image segmentation with color-alone feature using region growing pulse coupled neural network

Abstract Unsupervised color image segmentation based on low level color features aims to assign same label to all pixels of a region with color homogeneity, which underlies many higher level processing such as object detection and recognition. Pulse coupled neural network (PCNN) is a kind of biologically inspired spiking neural network, and has an inherent image segmentation nature which can combine each pixels intensity and its spatial relationship with neighboring pixels well. But PCNN cannot deal with color images, which restricts its applications greatly. For the problem, this article studied the color information embedding into PCNN and presented an unsupervised color image segmentation algorithm based on the proposed color region growing PCNN (CRG-PCNN) model. First, RGB color space is converted into Lab in which color distance can be evaluated linearly. Then a linking control unit (LCN) is introduced which essentially is a switch triggered by assessing the color distance among PCNN neuron feeding inputs. When the color distance between two pixels is less than a predefined threshold, a linking between the corresponding neurons is established. By doing so, color information can be embedded into PCNN effectively. Next, the L channel of an input image is fed into CRG-PCNN which can automatically pick out a seed neuron in each iteration and facilitates the region growing continuously by modifying the linking coefficient linearly with the assistance of a fast linking mechanism until certain termination conditions are met. Four widely used quantitative indices for massive experiments conducted on Berkeley segmentation dataset verify the performance of the proposed method. It has good segmentation accuracy and is concise and intuitive.

A parallel volume rendering method for massive data

Due to the limited memory capacity of current mainstream personal computer, increasing the volume rendering speed for massive 3D image data is a real challenge, especially for producing a high resolution output image. To achieve fast full-resolution volume rendering for the massive 3D data, a novel parallel volume rendering method based on ray casting is presented in this paper. Firstly, the original data is partitioned into sub-blocks. Then an occlusion relationship graph of all the sub-blocks is established, and a kind of topological sorting method is performed on this graph to find all the groups of sub-blocks without occlusion relationship. Finally, all the groups of sub-blocks are rendered by using a volume ray caster implemented on GPU, which renders the sub-blocks of each group in parallel. The experimental results indicate that the new method is effective and efficient for the volume visualization of massive 3D image data.