A novel filtering kernel based on difference of derivative Gaussians with applications to dynamic texture representation

Nguyen, Thanh Tuan; Nguyễn, Thanh Phương; Bouchara, Frédéric

doi:10.1016/j.image.2021.116394

Cited by 8 publications

(11 citation statements)

References 55 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…When using the 4CFV scheme, our DBRF correctly classifies all test samples, i.e., the rate is 100%. The rates provided by DFS [53], STRF N-jet [62], FoSIG [63], HoGF 2D [25], HoGF 3D [25], DoDGF 2D [65], MEMDP [76], RUBIG [74], and PI-LBP [15] are also 100%. Except for STRF N-jet and PI-LBP, all the other methods use SVM classifier.…”

Section: ) Results On the Ucla Datasetmentioning

confidence: 98%

“…The proposed DBRF has only 768 dimensions, which is substantially less than other methods. On the other hand, HoGF 2D [25], HoGF 3D [25], DoDGF 2D [65], DoDGF 3D [65], MEMDP [76], and RUBIG [74] slightly outperform DBRF by 0.5% and all of them adopt SVM for DT classification. Due to the fact that all of them use features from multiple scales, they generate highdimensional feature vectors (7200, 9600, 4800, 7200 3888, 21600, respectively), which are at least 4 times longer than ours.…”

Section: ) Results On the Ucla Datasetmentioning

confidence: 99%

“…Jansson and Lindeberg [62] used space-time separable kernels to extract spatio-temporal receptive field (STRF) responses, and then PCA is applied for dimension reduction. Nguyen et al [25], [63], [64], [65] incorporated completed LBP with Gaussian gradients, and reported a series of methods for DT representation.…”

Section: Filter-based Methodsmentioning

confidence: 99%

“…Due to the smoothing property of Gaussian kernel, the gradients are stable and robust against noise to some extent. Extracting gradients introduces extra directional information, which has been proved to be beneficial for DT presentation [65]. Second, random projection is conducted at every valid position in the three gradients, respectively.…”

Section: A Overview Of the Proposed Methodsmentioning

confidence: 99%

“…As for gradient computation, there exists at least five operators, i.e., Sobel operator, Prewitt operator, central difference operator, intermediate difference operator, and Gaussian first-order derivative. According to our previous experience on face recognition task [95] and the success of Nguyen et al [65], we adopt the Gaussian first-order derivatives for gradient extraction.…”

Section: B 3d Gaussian Gradients Extractionmentioning

confidence: 99%

See 4 more Smart Citations

Dynamic Texture Classification Using Directional Binarized Random Features

Zhao

et al. 2023

IEEE Access

View full text Add to dashboard Cite

Dynamic texture description has been studied extensively due to its wide applications in the field of computer vision. Local binary pattern (LBP) and its various variants account for a large part of dynamic texture description methods because of its advantages, such as good discriminability and low computational complexity. However, many LBP-based methods directly extract feature from pixel intensities and only use a proportion of pixels in a local neighborhood. And their good classification performance is usually achieved at the cost of high feature dimensionality, which would limit their application scenarios. We argue that extracting features from the gradient domain will capture more discriminative features due to the additional directional information, and that making use of all the pixels in a local neighborhood would improve performance. In this paper, we propose a simply but effective dynamic texture descriptor that inherits the advantages of LBP while excluding its disadvantages. The proposed method consists of four stages of data processing: 1) gradients extraction; 2) random feature extraction from gradients; 3) binary hashing of directional random features; and 4) histogramming. Gaussian first-order derivatives are used as gradient filters such that stable gradients could be generated. Then random projection is applied to extract random features from each gradients. Both the above two stages are conducted via 3D filtering, and thus they are efficient. Thirdly, the random features from each gradient are binarized and encoded into integer codes, from which a histogram is built. Finally, the histograms from each gradient are concatenated into a feature vector. Because we use 8-bit codes, The feature dimensionality is very low. We evaluate the proposed method on three benchmark dynamic texture datasets with various test protocols. The results demonstrate its effectiveness and efficiency when comparing to many state-of-the-art methods.

show abstract