Automatic classification of native wood charcoal
Introduction
Charcoal is an energetic source used in different segments of ironwork, metallurgic, cement industry, and others, being important economically and historically in Brazil. According to Davrieux et al. in (Davrieux et al., 2010), Brazil consumed more than 33 million cubical meters of charcoal, making it one of the biggest consumers as well as producers of this fuel in the world. Historically, in Brazil, wood from native forests was the main source used to produce charcoal, but today a combination of stricter environmental regulation and increased enforcement has increased the use of wood from planted forests (Yazdani et al., 2012). However, in certain regions, wood is still often carbonized from illegally cut trees. In such cases, enforcement of logging rules is hampered due to the difficulty of identifying the species, making it necessary to obtain more information on the intrinsic characteristics of carbonized wood to enable species identification (Nisgoski et al., 2014).
Usually, the identification of the charcoal is based on its anatomical characteristics which demand specialists in wood anatomy. However, few reach good accuracy in classification due to the time it takes for their training. To overcome such a limitation, some studies have been reported in the literature. The main line of investigation explores the fact that charcoal is a biocarbon produced by the carbonization of wood, a process which leads to a formation of a solid residue with an increased content of carbon element. With that in mind, the idea is to use a proper source of radiation to excite the wood surface to analyze the emitted spectrum. Near Infra-red spectroscopy (Labbé et al., 2006; Nisgoski et al., 2015; Ramalho et al., 2017) and Reflective spectroscopy in the Medium Infra-red region (Davrieux et al., 2010) are examples of the spectrum-based processing systems.
Another approach is the image-based, where some characteristics of the wood are extracted and analyzed to discriminate the species. Such a pattern recognition strategy has been successfully applied to forest species classification by several authors for both macroscopic and microscopic images (Kobayashi et al., 2015; Yadav et al., 2017; Yusof et al., 2013). After the release of public datasets of forest species (Martins et al., 2013; Paula Filho et al., 2014), some researchers have reported outstanding results using different textural descriptors and also deep learning techniques (Andrearczyk and Whelan, 2016; Hafemann et al., 2014).
However, the research on automatic classification of native wood charcoal is quite limited (Gonçalves and Scheel-Ybert, 2016; Nisgoski et al., 2014), mostly because of the lack of a robust public available dataset. To close this gap, in this work we introduce a wood charcoal database composed of 44 species. This database has been built in collaboration with the Laboratory of Wood Anatomy at the Federal University of Parana (UFPR) in Curitiba, Brazil, and it is available upon request for research purposes.1 The database introduced in this work makes future benchmark and evaluation possible.
In order to establish some baseline for further comparison, in this work we have assessed two configurations of the Local Binary Patterns (LBP) along with state-of-the-art machine learning classifiers. We also have evaluated representation learning using Convolutional Neural Networks. In our experiments we have observed similar results using handcrafted features and representation learning. Both representations achieved results around 95% of recognition rate.
The remaining of this paper is organized as follows: Section 2 presents the protocol used to built the proposed database of wood charcoal images. Section 3 describes the strategies used to represent the charcoal classification problem, where handcrafted and automatic representation were used. Section 4 presents the experimental protocol used to assess such different representations using the constructed wood charcoal database, and also the best results observed for each strategy. Finally, Section 5 presents our conclusions and future work.
Section snippets
Database
The wood charcoal database presented in this work contains 44 forest species with 12 images each as shown in Table 1. This balanced dataset were cataloged by the Laboratory of Wood Anatomy at the Federal University of Parana in Curitiba, Brazil. The trees were cut in a natural forest, and disks were extracted from the diameter at breast height (DBH), with a thickness of about 8 cm. Twelve samples were obtained from each species, with dimensions of 2 × 2 × 5 cm. Each sample was wrapped in
Features
In this section, we present the feature sets we have used to train the classifiers. Section 3.1 describes the handcrafted textural descriptors, while Section 3.2 gives the details about the CNN used to automatically extract the representation from the charcoal images.
Experimental results
In our experiments the dataset was randomly divided into training (50%) and testing (50%) (see Table 1). In addition, 20% of the training set was used for validation. Since the amount of images available in the dataset is limited, we have adopted the strategy presented by Hafemann et al. (Hafemann et al., 2014), where the original image is divided into several sub-images called patches. In this case, the main premise is that these patches can contain enough information to train a model, then an
Conclusion
In this paper we dealt with the challenging problem of automatic classification of native wood charcoal. For this purpose, we introduced a new database which is composed of images related to 44 different species of wood native charcoal. In addition, different methods were evaluated in the context of handcrafted and automatic feature learning. To assess the handcrafted features, we have used different state-of-the-art classifiers, however, experiments have shown that the best results were
Acknowledgements
This research has been supported by The National Council for Scientific and Technological Development (CNPq) grants 303513/2014-4 and 307277/2014-3.
References (27)
- et al.
Using filter banks in convolutional neural networks for texture classification
Pattern Recogn. Lett.
(2016) - et al.
Rlbp: Robust local binary pattern
- et al.
Survey on lbp based texture descriptors for image classification
Expert Syst. Appl.
(2012) - et al.
A comparative study of texture measures with classification based on featured distribution
Pattern Recogn.
(1996) - et al.
Investigation of the fuel value and the environmental impact of selected wood samples gathered from brunei darussalam
Renew. Sust. Energ. Rev.
(2012) - et al.
Discrimination of native wood charcoal by infrared spetrocopy
Quim Nova
(2010) - et al.
Charcoal anatomy of brazilian species. i. Anacardiaceae
Academina Brasileira de Ciências
(2016) - et al.
Forest species recognition using deep convolutional neural networks
Statistical and structural approaches to texture
Proc. IEEE
(1979)- et al.
Deep residual learning for image recognition
Automated recognition of wood used in traditional japanese sculptures by texture analysis of their low-resolution computed tomography data
J. Wood Sci.
Imagenet classification with deep convolutional neural networks
Evaluation of noise robustness for local binary pattern descriptors in texture classification
EURASIP Journal on Image and Video Processing
Cited by (14)
Fossil charcoal particle identification and classification by two convolutional neural networks
2019, Quaternary Science ReviewsCitation Excerpt :In palaeolimnology, Racca et al. (2003) use a multi-layer perceptron neural network to reduce diatom taxa for calibration purposes. Maruyama et al. (2018) use a convolutional neural network as a feature extractor to identify the species of native wood charcoal pieces, created from samples of modern trees. Weller et al. (2007) apply a supervised neural network to identification of sedimentary organic matter within pollen slides.
Preparation and application progress of photo-thermal conversion carbon materials for water treatment
2023, Gongneng Cailiao/Journal of Functional MaterialsComputer Vision-Based Wood Identification: A Review
2022, ForestsTexture Features and Statistical Features for Wood Types Classification System
2022, 2022 5th International Conference on Computer and Informatics Engineering, IC2IE 2022Automatic identification of charcoal origin based on deep learning
2021, Maderas: Ciencia y Tecnologia