Wood identification based on macroscopic images using deep and transfer learning approaches

Introduction

The identification and classification of wood species serve as crucial steps towards understanding their biological diversity, ecological role, economic value, and cultural significance. Accurate identification is also necessary for the appropriate use of wood (Wheeler & Baas, 1998). While identifying the species is relatively straightforward when organs such as flowers, leaves, or seeds are present, it becomes a challenging task once the wood is processed. The identification relies solely on the macroscopic or microscopic characteristics of the wood (Khalid et al., 2008). Macroscopic structure refers to the features observable to the naked eye or under low magnification, while microscopic structure denotes the cell-level characteristics observable under high magnification.

Sometimes, identification at the genus level is sufficient for diagnosing the type of wood. However, there are instances where species-level identification is necessary. For instance, while trade for certain wood species may be banned or restricted, other species within the same genus may be freely traded (such as Dalbergia spp., Pterocarpus spp.) (Gasson, 2011; Shou et al., 2014; Snel et al., 2018). Furthermore, species-level identification is needed when gathering information about structures of cultural and historical value (Hwang et al., 2016). Microscopic methods are preferred in such cases; however, some wood species cannot be distinguished from each other microscopically (Tuncer, 2020). Therefore, the work on improving methods for diagnosing wood species continues, with image processing techniques showing significant advancements due to emerging technologies.

Traditional methods for diagnosing the type of wood involve examining its macroscopic or microscopic structure. The microscopic technique examines the wood’s anatomy using light or electron microscopy. It identifies cell types, measures cell wall thickness, determines cell sizes and shapes, analyzes cell contents, and examines the wood’s cell wall chemistry. The macroscopic method observes the wood’s physical and anatomical features with the naked eye or a hand magnifier. It identifies properties such as color, density, hardness, odor, grain, fibers, pores, luster, texture, resin, grain structure, cracks, knots, and more. This reveals the wood’s natural and unique qualities, which can vary depending on factors like age, growth conditions, cutting time, drying method, processing, decay level, and staining or varnishing. Although these methods are considered the most reliable, they are time-consuming, labor-intensive, and require expertise in wood anatomy. As a result, there is a growing demand for new and different diagnostic methods (Dormontt et al., 2015). In an era where quick and easy access to information is expected, new technologies need to be developed for industries working with wood materials and for maintaining environmental balance.

The limitations of these traditional methods underscore the need for more practical, faster, and cheaper alternatives in diagnosing wood types. One such alternative involves computer vision techniques compatible with widely used devices like smartphones. Consequently, various studies have been conducted in the area of wood type recognition and classification using computer vision techniques (Martins et al., 2013; Yadav et al., 2013, 2015; Rosa Da Silva et al., 2017; Fabijańska, Danek & Barniak, 2021). These studies have used macroscopic or microscopic images of wood samples. Macroscopic images are 10–15× magnified photographs of wood surfaces, while microscopic images are photographs of wood samples magnified between 25–100× under a microscope. The species of wood are identified by applying feature extraction and classification processes on these images. The drawback of these methods, however, is the need for specialized equipment to capture the images (Filho et al., 2014).

In the process of wood species recognition through image processing, a two-stage model is typically employed: feature extraction and classification. During the feature extraction phase, a variety of methods can be utilized to convert images into numerical data. One such method is the Gray Level Co-occurrence Matrices (GLCM). This technique calculates pixel gray level differences that reflect the texture properties of the image. The features derived from the image are then used as input in the classification stage (Khalid et al., 2008; Piuri & Scotti, 2010; Mallik et al., 2011; Martins et al., 2013; Yadav et al., 2013).

In the classification phase, wood samples are trained using different algorithms, and a statistical model is built. This model classifies new examples according to wood species. Among these algorithms are multilayer artificial neural networks (Yusof & Rosli, 2013; Zhao, Dou & Chen, 2014; Ibrahim et al., 2018) and support vector machines (Turhan & Serdar, 2013; Martins et al., 2013; Filho et al., 2014; Barmpoutis et al., 2018), and methods like KNN. These methods have their pros and cons, and their performance depends on the dataset and feature extraction method (Fabijańska, Danek & Barniak, 2021).

However, traditional methods for identifying and classifying wood species often require examination of the wood’s anatomical structure, which can be time-consuming and require expertise. Deep learning, on the other hand, employs artificial neural networks that learn from and make predictions based on data. Thus, they can analyze data, learn from it, and make predictions. As a result, deep learning techniques are yielding faster and more effective results in this field. These techniques can also learn to identify new types of wood. The model can be updated with new data and cover a broader range of wood species. However, deep learning techniques require a substantial amount of data for accurate wood type diagnosis. To make accurate predictions, images of a wide variety of wood types are needed. Gathering and labeling this data can be challenging. Deep learning techniques can be affected by factors like image quality, lighting, angle, background, etc. These factors can lead to image degradation or misinterpretation.

Sun et al. (2021) developed a transfer learning approach to wood species identification using a limited data set. Linear discriminant analysis (LDA) and k-nearest neighbours (KNN) were used as classification methods to enhance the features obtained from the ResNet50 model (Sun et al., 2021).

However, Fabijańska, Danek & Barniak (2021), using a macroscopic data set of 14 species of corals and angiosperms common in Europe, developed a method that now uses an evolutionary encoder network and a sliding window layout. This study shows that a method that now uses a sliding window layout with an evolutionary encoder network can diagnose wood species with high accuracy, regardless of the spatial resolution of the images.

A study by Kırbaş & Çifci (2022) compared the transfer learning performance of various deep learning architectures on a WOOD-AUTH dataset comprising 12 wood species. The study evaluated the transfer learning performance of ResNet50, Inceptionv3, Xception, and VGG19. The results show that the Xception model achieves the highest accuracy (Kırbaş & Çifci, 2022). In a related study, researchers proposed a wood recognition approach utilizing ResNet50 to extract the textile properties of wood, applying global average pooling (GAP) to decrease attribute numbers and boost model generalizability. Extreme learning machine (ELM) algorithm was then implemented for classification. ELM is a one-layer advanced neural network capable of learning properties from images without adjusting the parameters of hidden layer or backward spread.

Elmas (2021) applied pre-trained convolutional neural networks such as AlexNet, DenseNet201, ResNet18, ResNet50, ResNet101, VGG16, and VGG19 on a data set consisting of 24,686 wood bark images belonging to 59 wood species collected from various regions of Turkey using the transfer learning method, and stated that the highest accuracy value was obtained with DenseNet121 (Elmas, 2021). Miao et al. (2022) proposed a convolutional neural network developed for wood type recognition from wood images in their study. The model combines Inception and mobilenetV3 networks. The model was tested with a large-scale data set containing 16 different wood types. The model achieved 96.4% accuracy (Miao et al., 2022). Wu et al. (2021) used deep learning methods for wood identification based on longitudinal cross-section images in their study. The model benefited from effective convolutional neural networks such as ResNet50, DenseNet121 and MobileNetv2. The model was evaluated with a data set containing 11 common wood types. The model achieved 98.2% accuracy (Wu et al., 2021). These studies reveal that deep learning techniques are more successful than traditional methods in the field of wood type identification and classification.

There are some specific challenges in identifying wood species. The anatomical structure of the species, the prevailing climate and soil conditions, forest density in the region, the amount and direction of sunlight received, among other factors, can affect the characteristics and classification of wood species. Another challenge is that large, public data sets related to wood images are only available for a specific region. The transfer learning method allows pre-trained convolutional neural networks to be used to extract features from wood images. This reduces the need for more data to train the model and increases the model’s generalization ability.

In this study, a transfer learning method was used that fine-tuned pre-trained convolutional neural networks (CNN) models such as ResNet50, DenseNet121, MobileNetv2 and ShuffleNetv2. This method transfers the information learned from ImageNet, a comprehensive data set consisting of natural images, to our task of determining wood types using macroscopic images. Most of the state-of-the-art models examined did not report on different performance metrics that help gain more insight into the proposed model and solution approach. The structure of CNN models is becoming increasingly complex, hardware costs are rising and application efficiency is decreasing, therefore lightweight deep learning model research is an inevitable trend. An attempt was made to achieve a success close to more complex models by using a customized ShuffleNet architecture with different parameters in the recognition of wood species.

Materials and Methods

Recognition of wood species typically demands a cross-sectional review of wood features by experts. These characteristics include vessel arrangement, pore direction, ray parenchyma, fiber, and other surface information. The identification standards from the IAWA list (Wheeler & Baas, 1998) are used to compare these features between wood species (Sun et al., 2021).

Dataset

A dataset of macroscopic images was used for forest species detection and classification (Cano Saenz et al., 2022). Ten species were chosen for analysis, ensuring a minimum of 800 images for each category. In total, the dataset comprised 10,792 images of tropical forest species. The training module employed a subset of 8,000 images chosen at random, while the test group incorporated a maximum of 2,221 images. This equal allocation of the training data will facilitate a more precise assessment of the model’s overall performance. Detailed information regarding the types employed and the total number of images can be found in Table 1.

Table 1:

The species and image numbers used in this study.

	Tür	Train	Test	Image number
1	Achapo	800	389	1,189
2	Cedro costeno	800	328	1,128
3	Chanul	800	201	1,001
4	Cipres	800	15	815
5	Cuangare	800	300	1,100
6	Eucalipto blanco	800	305	1,105
7	Guayacan amarillo	800	306	1,106
8	Nogal cafetero	800	129	929
9	Sajo	800	23	823
10	Urapan	800	225	1,025
	Total	8,000	2,221	10,792

DOI: 10.7717/peerj.17021/table-1

The data set contains macroscopic images obtained using a scale that allows the observation of pores, fibers and parenchyma. The data set includes more than 8,000 images with a resolution of 640 × 480 pixels, containing 3.9 microns per pixel, and an area of (2.5 × 1.9) square millimeters where anatomical features are displayed. Figure 1 shows macroscopic images of species.

Experimental results

In this study, RestNet18, GoogLeNet, VGG19, Inceptionv3, MobileNetv2, DenseNet201, InceptionResNetv2, EfficientNetb0, ShuffleNet deep learning models were used for transfer learning. These models were trained for a general-purpose visual recognition task and learned high-level features to recognize different objects. The last layer of these models was removed and a new layer with neurons equal to the number of forest species was added. This new layer was initialized with random weights and trained with the forest species data set. Thus, the model adapted the general features learned in the source domain to the target domain.

The models in Table 2 are artificial neural networks that are used for visual processing tasks such as image classification. The size and number of parameters of each model indicate the complexity and efficiency of the model. The number of parameters specifies the number of weights that the model can learn.

The classification accuracies of the fine-tuned models on the test images were calculated (Table 3). No layer was frozen during the model execution. Data augmentation was not applied. The mini batch size was set to 64. The learning rate was chosen low because transfer learning does not require training for many epochs. Retraining updates the network to learn and identify features associated with new images and categories. In most cases, retraining requires less data than training a model from scratch. After the model was retrained, a total of 2,221 test images were classified and the performance of the networks was evaluated.

Different metrics can be used to compare the performance of image classification models. To determine the model that shows the best performance according to any of these metrics, it is necessary to train and test different models on the same data set. In this way, it is possible to compare the results of the models objectively. All these parameter values are obtained for the same database and compared.

The modified ShuffleNet requires more training than other models because some weights and biases are not transferred. Table 1 shows the effect of the changes made on the model performance. Figure 2 presents the confusion matrix for the test images of the ShuffleNet model that was changed with the proposed method. The confusion matrix is a table that summarizes the prediction results for a classification problem. The confusion matrix allows us to get an idea about the types that are misclassified by the classifier. The total accuracy of the model is 92%, which means that it correctly predicts 92% of the test data.

The model has a high accuracy for classes 1, 2, 3, 6 and 7, which are Achapo, Cedro costeno, Chanul, Eucalipto Blanco and Guayacan Amarillo. The accuracy of these classes is above 95% and there are very few misclassifications. The model has a low accuracy for class 4, which is Cipres. This class has only 71% accuracy and many misclassifications. Species-level identification of wood types is more difficult than genus-level, because some wood types within the same genus can have very similar macroscopic features that are hard to distinguish by human eye or computer vision techniques (Schweingruber & Baas, 2011).

Discussion

Several pre-trained CNN models such as ResNet50, DenseNet121, MobileNetv2 and ShuffleNetv2 were fine-tuned using a transfer learning method in this study. This method enables the transfer of the knowledge learned from ImageNet, which is a large and diverse data set of natural images, to the task of identifying wood types using macroscopic images (Abu Al-Haija, 2022). Other transfer learning methods that can be applied for wood species identification are feature extraction, where pre-trained models are used as feature extractors and a new classifier is trained on top of them; or multi-task learning, where the pre-trained models are trained on multiple related tasks simultaneously (Elmas, 2021). The weights of the previous layers in the network can be optionally frozen by setting the learning rates in those layers to zero (Katsigiannis et al., 2023). The parameters of the frozen layers are not updated during training, which can significantly speed up the network training. If the new data set is small, freezing weights can also prevent the network from overfitting to the new data set. In general, if the data set is small or similar, fewer layers are frozen or fewer weights are changed. If the data set is large or different, more layers are frozen or more weights are changed. If the target data set is large, this means that there is enough data to learn and avoid overfitting. Therefore, freezing more layers can speed up the training process and reduce the computation cost without sacrificing much accuracy. It can also prevent the model from forgetting the useful features learned from the source data set; changing more weights can help the model adapt to the new domain (Sanida et al., 2023).

There are many pre-trained models available for transfer learning, and each has its advantages and disadvantages that need to be considered: When choosing these models, factors such as size, accuracy and prediction speed should be taken into account. Size, the importance of your model size will vary depending on where and how you plan to use it. Accuracy, shows the generalization ability of the model. However, a low accuracy score on ImageNet does not mean that the model will perform poorly on all tasks. Prediction speed determines the performance of the model. These factors can vary depending on the architecture of the model and the data set it was trained on. The model structure of CNNs is becoming more and more complex, the hardware cost is increasing and the application efficiency is decreasing, so lightweight deep learning model research is an inevitable trend (Song, Zhang & Long, 2023).

Lightweight and computationally efficient models; MobileNetv2 and ShuffleNet are good options when there are limitations on the model size. MobileNetV2 is good for reducing network size and cost but contains subsampling bottlenecks (Pradipkumar & Alagumalai, 2023). ShuffleNetv2 is a convolutional neural network optimized for speed and memory. ShuffleNet has been reported to achieve ~13X real speed increase compared to AlexNet CNN while maintaining similar accuracy levels (Abu Al-Haija, 2022). There are studies that place the compression and excitation blocks after each stage of the Shuffle Block to reduce the computation cost. New architectures based on ShuffleNets are being designed (Tang et al., 2020; Pradipkumar & Alagumalai, 2023).

ShuffleNet is a good choice for wood species identification because it is designed to reduce the computational complexity and increase the efficiency of convolutional neural networks. It uses two techniques called group convolution and channel shuffle to achieve this goal. These techniques reduce the number of parameters and operations while increasing the feature diversity and information flow of the model. ShuffleNet has potential advantages over other models such as mobileNetv2 for wood species identification. ShuffleNet is faster and more accurate than MobileNetv2 in various tasks such as image classification, object detection and face recognition (Zhang et al., 2018). Moreover, as our experimental results also show, ShuffleNet can adapt better to different data sets and transfer learning methods.

The limitations and challenges of this study are that collecting, processing and labelling wood images is a laborious and time-consuming process. Also, situations such as the same wood species looking different in different regions or different wood species resembling each other can make identification difficult. There is no standard data set or evaluation metric for wood species identification with deep learning. Therefore, it is hard to compare or generalize the results of different studies. Most of the studies have used regional data bases and focused on a limited number of species. There is a need for a more general reference database (Ergun & Uzun, 2022).

Conclusions

In this study, different deep learning architectures were used to identify wood species. The architectures used include ResNet, Inceptionv3, Xception, VGGNet, GoogLeNet, MobileNetv2, DenseNet201 and ShuffleNet. Metrics such as accuracy, sensitivity, precision, F1-score, specificity and MCC were used to evaluate the performance. Deeper and more complex models performed better. However, these models required more computation power and time. The modified ShuffleNet reduces complexity and computation costs while achieving the same level of success.

Machine and deep learning are powerful techniques that can be used to identify wood species. However, these techniques need lightweight networks to work on mobile devices. The results show that the modified ShuffleNet model achieves similar performance to more complex models while requiring less computation power, making it suitable for mobile applications. The study demonstrates the potential of deep learning techniques for identifying wood types using macroscopic images. It also provides a benchmark for future research in this field. In the next studies, feature extraction and transfer learning methods can be compared; different types of data sets can be used, the degrees of similarity or difference between different wood species can be examined; a more general reference data set can be created; more user-friendly and accessible applications and tools can be designed for wood species identification.

In future studies, it may be possible to improve the performance of wood species identification by using different feature extraction methods and different transfer learning methods (multi-task learning). Also, increasing the generalizability and reliability of wood species identification by using larger, more diverse and quality data sets or creating synthetic data sets is an important research topic.

Supplemental Information