HeadPose-Softmax: Learning loss of head pose adaptive curriculum for deep face recognition (2023)

Expert Systems with Applications, Volume 51, 2016, pp. 207-217

To learn a Bayesian network classifier, continuous attributes usually need to be discretized. But discretizing continuous attributes can lead to missing information, noise, and less sensitivity to the change of attributes versus class variables. In this paper, we use the Gaussian kernel function with smoothing parameters to estimate the density of attributes. The Bayesian network classifier with continuous attributes is built by dependency extension of naive Bayes classifiers. We also analyze the information provided to a class for each attribute as a basis for dependency extension of naive Bayes classifiers. Experimental studies on UCI datasets show that Bayesian network classifiers using the Gaussian kernel function provide good classification accuracy when dealing with continuous attributes compared to other approaches.

Pattern Recognition, Volume 138, 2023, Item 109336

Efficient detectors of rotationally symmetric shapes are proposed by introducing a novel concept of half-shape signatures to overcome the main problem of projection-based approaches to study the rotationally symmetric properties of any binary shape. In fact, the fact is that the projection cues in these conventional approaches are periodic with a period of$\mathrm{Pi}$has limited an applicable use of rotational symmetry detection. To this end, we propose a new concept of the profile of half-shapes as a shape signature, together with a simple but efficient technique such that the rotational symmetry of the binary shape can be determined by considering the correlation between this signature and its circular displacement. In addition, a new meaningful measure is also introduced, ranging from 0 to 1 to indicate how perfect the rotational symmetry would be. Experimental results to detect single/composite shapes have clearly confirmed the competence of our proposal.

Pattern Recognition, Volume 115, 2021, Item 107895

Dynamic time warp ($\text{DTW}$) is a popular similarity measure for matching and comparing time series. Because of$\text{DTW}$Due to the high computational time of , lower bounds are often used to check for bad matches. Many alternative lower bounds have been proposed, offering a range of different tradeoffs between tightness and computational efficiency.$\text{LB}\_\text{KEOGH}$offers a useful compromise in many applications. Two current lower bounds,$\text{LB}\_\text{IMPROVED}$And$\text{LB}\_\text{ADVANCED},$are much narrower than$\text{LB}\_\text{KEOGH}$. All three have the same worst case computational complexity - linear with respect to row length and constant with respect to window size. We introduce four new ones$\text{DTW}$Lower bounds in the same complexity class.$\text{LB}\_\text{SMALL JEANS}$is much tighter than$\text{LB}\_\text{IMPROVED},$with only modest additional computational effort.$\text{LB}\_\text{WEBB}$is more efficient than$\text{LB}\_\text{IMPROVED},$while often providing a closer bond.$\text{LB}\_\text{WEBB}$is always tighter than$\text{LB}\_\text{KEOGH}$. The parameter free$\text{LB}\_\text{WEBB}$is usually tighter than$\text{LB}\_\text{ADVANCED}$. A parameterized variant,LB_Webb_Enhanced, is always narrower than$\text{LB}\_\text{ADVANCED}$. Another variant$\text{LB}\_{\text{WEBB}}^{*},$is useful for some restricted distance functions. In extensive experiments$\text{LB}\_\text{WEBB}$proves to be very effective for finding nearest neighbors.

Pattern Recognition, Volume 139, 2023, Item 109507

This work focuses on the problematic of personalized federated learning (FL) with contrastive learning (CL) scheme, which aims to implement a collaborative pattern classification by many clients. The traditional FL frameworks mostly allow the global model for the server and the local models for the clients to be similar, often ignoring the data heterogeneity of the clients. Aiming at achieving better performance in clients, this study introduces a personalized federated contrastive learning model called PerFCL by proposing a new approach to double contrastive representation learning (DCL). Specifically, PerFCL borrows the DCL scheme where one CL loss compares the shared parts of local models to the global model and the other CL loss compares the personalized parts of local models to the global model. To promote the difference between the two parts, we created a double optimization problem consisting of maximizing comparison agreement for the former and minimizing comparison agreement for the latter. We evaluated the proposed model against three publicly available datasets for federated image classification. The trial results show that PerFCL benefits from the proposed DCL strategy and outperforms the state-of-the-art federated learning models.

Journal of Visual Communication and Imaging, Volume 59, 2019, pp. 195-203

Research has shown that examining an image holistically provides a better understanding of the image than separate processes, each dedicated to a single task such as annotation, classification, or segmentation. During the last decades there have been several efforts for simultaneous image classification and annotation using probabilistic or neural network based topic models. Despite their relative success, most of these models suffer from poor visual word representation and imbalance between the number of visual and annotation words in the training data. This paper proposes a novel simultaneous image classification and annotation model based on SupDocNADE, a neural network-based theme model for image classification and annotation. The proposed model, named wSupDocNADE, addresses the above shortcomings by using a new encoding and introducing a weighting mechanism for the SupDocNADE model. In the coding step of the model, several patches extracted from the input image are first fed into a deep convolutional neural network and the feature vectors obtained from this network are coded using LLC coding. These vectors are then aggregated into a final descriptor by sum pooling. To overcome the imbalance between the figurative and annotation words, a weighting factor is considered for each figurative or annotation word. The weights of the visual words are determined based on their frequencies obtained from the pooling method, and the weights of the annotation words are learned from the training data. The experimental results on three benchmark datasets demonstrate the superiority of the proposed model over state-of-the-art models in both image classification and annotation tasks.

Pattern Recognition Letters, Band 128, 2019, S. 107-114

Most existing machine learning algorithms do not natively support nominal attributes, so it is important to develop the data transformation from nominal attributes to high quality numeric attributes. The conditional probability transformation (CPT), which uses conditional probability terms to transform categories into nominal attributes, is competitive with modern transformation methods such as One-Hot Encoding (OHE) and Separability Split Value Transformation (SSVT). However, conditional probability terms can be difficult to estimate accurately when the training data is insufficient or when there are strong dependencies between their attributes. Inspired by the fine-tuning method for improving conditional probability terms in distance measures, we proposed a fine-tuned conditional probability transformation (FTCPT). Furthermore, we proposed an improved SSV (ISSV) based on fine-tuned conditional probability terms and used our modified MIC-based feature selection method to further improve the performance of FTCPT. Experimental results show that the proposed methods can improve the quality of the data transformation and thereby help improve the classification performance of the subsequent machine learning algorithm.