Multi-Level Texture Encoding and Representation (Multer) Based on Deep Neural Networks

Hu, Yuting; Long, Zhiling; AlRegib, Ghassan

doi:10.1109/icip.2019.8803640

Cited by 20 publications

(29 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For the first component, we investigated the effects of a) normalization , b) sum-to-one constraint across bins, and c) location of histogram layer. Previous works incorporated batch normalization of their convolutional features before fusing their texture features (i.e., concatenation, weighting) [20,43,46]. We wanted to investigate whether the impact of batch normalization for our new layer would improve results similar to existing "encoding layers."…”

Section: Methodsmentioning

confidence: 99%

“…Previous histogram-based features used lower level features (e.g., edges [16] for edge histogram descriptors) and learning histograms of convolutional features earlier in the network may exploit more texture information. Training Details A similar training procedure from [20,43] was used in this work. For each dataset, the image is resized to 256×256 and a random crop of 80 to 100 % of the image was extracted with a random aspect ratio of 3/4 to 4/3.…”

Section: Methodsmentioning

confidence: 99%

“…After the best configuration of the HistRes model was determined, the performance was compared using overall acccuracy between different "encoding layers" for each version of ResNet: 1) global average pooling (GAP), 2) Deep Texture Encoding Network (DeepTEN) [46], 3) Deep Encoding Pooling (DEP) [43], 4) Fisher Vector CNN (FV-CNN) [13] 5) Multi-level texture encoding and representation (MuLTER) [20] and 6) our HistRes B. We also perform T-Tests to analyze the significance of the difference between results of comparison methods.…”

Section: Methodsmentioning

confidence: 99%

“…The DEP network also retains both spatial and texture information, but each HistRes B achieved slightly better accuracy than the DEP model. The MuLTER [20] incorporates texture at multiple levels and the batch size for MuLTER was set to 32, resulting in the model seeing the data more times (in comparison to our batch size of 128). Despite these differences, our method achieves slightly better accuracy as well and there is no significant difference between the results obtained.…”

Section: Comparison To Other Texture Encoding Approachesmentioning

confidence: 99%

See 3 more Smart Citations

Histogram Layers for Texture Analysis

Peeples,

Xu,

Zare

2020

Preprint

View full text Add to dashboard Cite

We present a histogram layer for artificial neural networks (ANNs). An essential aspect of texture analysis is the extraction of features that describe the distribution of values in local spatial regions. The proposed histogram layer directly computes the spatial distribution of features for texture analysis and parameters for the layer are estimated during backpropagation. We compare our method with state-of-the-art texture encoding methods such as the Deep Encoding Network Pooling (DEP) [43], Deep Texture Encoding Network (DeepTEN) [46], Fisher Vector convolutional neural network (FV-CNN) [13], and Multi-level Texture Encoding and Representation (MuLTER) [20] on three material/texture datasets: (1) the Describable Texture Dataset (DTD) [12]; (2) an extension of the ground terrain in outdoor scenes (GTOS-mobile) [43]; (3) and a subset of the Materials in Context (MINC-2500) dataset [6]. Results indicate that the inclusion of the proposed histogram layer improves performance. The source code for the histogram layer is publicly available 1 .

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Comparison To Other Texture Encoding Approachesmentioning

confidence: 99%

See 2 more Smart Citations

Histogram Layers for Texture Analysis

Peeples,

Xu,

Zare

2020

Preprint

View full text Add to dashboard Cite

show abstract

“…To combine both low-and high-level CNN features, we propose a multi-level texture encoding and representation network (MuLTER) on top of our previous work [49], whose architecture is shown in Fig. 6 and Table II.…”

Section: Proposed Learning-based Methods For Materials Surface Charac...mentioning

confidence: 99%