Crack Severity Classification from Timber Cross-Sectional Images Using Convolutional Neural Network

Featured Application

Application of convolutional neural network (CNN) for automatic evaluation of internal crack severity in building square timbers. In our previous article, we constructed a simple CNN and employed cross-sectional pictures covered with silver paint to make the cracks easier to recognize. However, in this article, we confirmed that the severity of the internal cracks could be evaluated with high accuracy by a ResNet50-CNN (Residual Neural Network 50 CNN), even in the case of unpainted images. Given that quality managers must evaluate numerous timbers, this automation process reduces the workload.

Abstract

Cedar and cypress used for wooden construction have high moisture content after harvesting. To be used as building materials, they must undergo high-temperature drying. However, this process causes internal cracks that are invisible on the outer surface. These defects are serious because they reduce the strength of the timber, i.e., the buckling strength and joint durability. Therefore, the severity of internal cracks should be evaluated. A square timber was cut at an arbitrary position and assessed based on the length, thickness, and shape of the cracks in the cross-section; however, this process is time-consuming and labor-intensive. Therefore, we used a convolutional neural network (CNN) to automatically evaluate the severity of cracks from cross-sectional timber images. Previously, we used silver-painted images of cross-sections so that the cracks are easier to observe; however, this task was burdensome. Hence, in this study, we attempted to classify crack severity using ResNet (Residual Neural Network) from unpainted images. First, ResNet50 was employed and trained with supervised data to classify the crack severity level. The classification accuracy was then evaluated using test images (not used for training) and reached 86.67%. In conclusion, we confirmed that the proposed CNN could evaluate cross-sectional cracks on behalf of humans.

1. Introduction

Cedar and cypress used for wooden construction have high moisture content after harvesting. To be used as building materials, they must undergo high-temperature drying [1]. This process is indispensable to prevent deformation and mildew. However, if the timbers are not dried using the appropriate procedure, internal cracks are formed, as shown in Figure 1a [2].

Internal cracks are severe because they reduce the durability of timber [3,4]. Based on this knowledge, physical stress propagation during drying was analyzed to improve the high-temperature drying technique with less cracking [5,6]. In the actual field, severe internal cracks can occur because of excessive high-temperature drying.

As shown in Figure 2a, the quality manager extracted a sample block of timber. In Figure 2b, the quality manager had to decide whether to use the material for the building based on the severity of the cracks by measuring the crack length, width, and number using calipers. However, this task is labor-intensive.

Therefore, this study aims to realize an intelligent system that automatically evaluates the severity of internal cracks based on cross-sectional images using a convolutional neural network (CNN) [7], as shown in Figure 2c.

We developed a simple CNN dealing with silver-painted cross-sectional images to make the cracks more clearly visible, as shown in Figure 3a [8]. The simple CNN that we proposed in our previous article could evaluate crack severity with approximately 85.67% accuracy [9]. However, the silver-painting task is cumbersome. Thus, in the present study, the proposed CNN deals with unpainted images, as shown in Figure 3b.

In previous studies related to the internal cracks in the timber, we found attempts to infer the internal cracks using acoustic signals [10,11,12]. However, our objective differs compared to these studies because they attempt to infer the condition of invisible internal cracks. In contrast, our attempt is to evaluate the visible cross-section depending on the crack severity.

Other studies applied CNNs to automatically detect cracks on wooden building surfaces [13] in order to find critical damage in wooden buildings. We also found studies to find cracks, knots, and mold in timbers [14,15] using CNN. Meanwhile, our study differs because we aimed to automatically evaluate crack severity levels rather than to detect cracks. In other words, most of the relevant studies focused on the automatic detection of cracks and not on the evaluation of the level of severity in the detected cracks. The present paper differs from other similar studies in that it focuses on evaluating the rank of the timber’s quality. We have not found any study that attempts to employ a CNN to evaluate the quality of square timber considering cracks.

Quality control staff in vendors, and researchers improving high-temperature drying techniques, have to rank many square timbers visually, and the burden of this task is considerable. The practical application of this study is to replace humans with artificial intelligence that automatically evaluates the quality of timbers.

Our previous study [8,9] constructed a simple CNN to evaluate silver-painted cross-sectional grayscale images. In the painted image, only the cracks are visible, excluding annual rings or knots. However, the painting process is labor-intensive. Therefore, the present paper targets color images without silver paint, and we employed a CNN taking color images as the input.

Convolutional neural networks are more suitable for capturing image features than a classical fully connected neural network because the CNN focuses on the relations of neighboring pixels of an image. On the other hand, the classical fully connected neural networks are inefficient in learning because they consider relationships between all pixels with the same importance.

The remainder of this paper is organized as follows. Section 2 describes the method for preparing and classifying timber cross-sectional images and presents the proposed CNN structure. Section 3 explains the validation method and results for the proposed CNN. Section 4 discusses the given result and future perspectives. Section 5 describes the conclusion of the present paper.

2. Materials and Methods

This section describes procedures to prepare cross-sectional images of timbers. In addition, the CNN configuration to evaluate the crack severity of the cross-sectional image is explained.

2.1. Timber Image Acquisition and Crack Severity Classification

We prepared 32 square cedar timbers with a 3000 mm length, as shown in Figure 4a, and then carried out high-temperature drying. Orange block pieces were cut out, and 32 blocks were obtained for the upper and root sides. As shown in Figure 4b, we scanned the images of the cross-sections surrounded by the green dashed lines using a scanner. Hence, 32 pictures for the upper side blocks and 32 pictures for the root side blocks were obtained, i.e., a total of 64 pictures.

A crosscut saw (KM-3C; Kobayashi Kikai Kogyo Co., Ltd.; Mishima-City, Shizuoka, Japan) was used to cut square timbers, as shown in Figure 5a. A belt sander (HUS-3: Hasegawa Wood Tech Co., Ltd.; Fuchu-City, Hiroshima, Japan) was used to flatten the cross-section. A scanner (Canon CanoScan LiDe 700F) was used to photograph the cross-section, as shown in Figure 5b. The scanning resolution was 200 dpi.

We classified the obtained 64 images depending on the cracking severity into classes A (Good), B (Neutral), and C (Bad), expressing the timber condition. Examples of these images are shown in Figure 6. Figure 6a shows the images categorized as class A (Good) with almost no cracks. Figure 6b shows class B (Neutral) with minor cracking. Figure 6c, which is labeled as class C (Bad), shows a lot of noticeable cracking with risk to strength. Among all 64 images, 20 images were in class A (Good), 20 in class B (Neutral), and 24 in class C (Bad).

2.2. CNN Structure for Crack Severity Evaluation

The input to the proposed CNN was a 244 × 244 × 3 RGB (Red-Green-Blue color model) image, as shown in Figure 7. The outputs are classification probabilities in the range of 0 to 1, expressed in terms of classes A (Good), B (Neutral), and C (Bad), representing the condition of the timber. The class with the highest probability is the classification result.

The proposed CNN is based on ResNet50-CNN (Residual Neural Network 50 CNN) [16]. However, the CNN is not pre-trained with image data, such as ImageNet [17]. Initially, the convolution filters and connection weights were random. Although ResNet50-CNN has very deep layers, it uses a skip structure and 1 × 1 convolution operation to improve learning efficiency. We also found that ResNet-CNN can be applied to the feature extraction of cracks in concrete infrastructure, such as roads or bridges [18].

Table 1. CNN feedforward calculation from input to stack 1.

Layer		Operation	Filters’ Set Num	Filter Size	Stride, Zero-Padding [Top, Bottom, Left, Right]	Feature Map Output
Input		-	-	-	-	224 × 224 × 3
Conv 1		Conv, BatchNorm, ReLU	64	7 × 7 × 3	2 × 2, [2,3,2,3]	112 × 112 × 64
Max pool		Max pooling	-	3 × 3	2 × 2, [0,1,0,1]	56 × 56 × 64
Stack 1_1	Conv 1_1_1	Conv, BatchNorm, ReLU	64	1 × 1 × 64	1 × 1, [0,0,0,0]	56 × 56 × 64
	Conv 1_1_2	Conv, BatchNorm, ReLU	64	3 × 3 × 64	1 × 1, [1,1,1,1]	56 × 56 × 64
	Conv 1_1_3	Conv, BatchNorm	256	1 × 1 × 64	1 × 1, [0,0,0,0]	56 × 56 × 256
Stack 1_1_br	Conv 1_1_4	Conv, BatchNorm	256	1 × 1 × 64	1 × 1, [0,0,0,0]
Stack 1_1_add		Add, ReLU	-	-	-
Stack 1_2	Conv 1_2_1	Conv, BatchNorm, ReLU	64	1 × 1 × 256	1 × 1, [0,0,0,0]	56 × 56 × 64
	Conv 1_2_2	Conv, BatchNorm, ReLU	64	3 × 3 × 64	1 × 1, [1,1,1,1]	56 × 56 × 64
	Conv 1_2_3	Conv, BatchNorm	256	1 × 1 × 64	1 × 1, [0,0,0,0]	56 × 56 × 256
Stack 1_2_br	-	-	-	-	-
Stack 1_2_add		Add, ReLU	-	-	-
Stack 1_3	Conv 1_3_1	Conv, BatchNorm, ReLU	64	1 × 1 × 256	1 × 1, [0,0,0,0]	56 × 56 × 64
	Conv 1_3_2	Conv, BatchNorm, ReLU	64	3 × 3 × 64	1 × 1, [1,1,1,1]	56 × 56 × 64
	Conv 1_3_3	Conv, BatchNorm	256	1 × 1 × 64	1 × 1, [0,0,0,0]	56 × 56 × 256
Stack 1_3_br	-	-	-	-	-
Stack 1_3_add		Add, ReLU	-	-	-

,

Table 2. CNN feedforward calculation in stack 2.

Layer		Operation	Filters’ Set Num	Filter Size	Stride, Zero-Padding [Top, Bottom, Left, Right]	Feature Map Output
Stack 2_1	Conv 2_1_1	Conv, BatchNorm, ReLU	128	1 × 1 × 256	2 × 2, [0,0,0,0]	28 × 28 × 128
	Conv 2_1_2	Conv, BatchNorm, ReLU	128	3 × 3 × 128	1 × 1, [1,1,1,1]	28 × 28 × 128
	Conv 2_1_3	Conv, BatchNorm	512	1 × 1 × 128	1 × 1, [0,0,0,0]	28 × 28 × 512
Stack 2_1_br	Conv 2_1_4	Conv, BatchNorm	512	1 × 1 × 256	2 × 2, [0,0,0,0]
Stack 2_1_add		Add, ReLU	-	-	-
Stack 2_2	Conv 2_2_1	Conv, BatchNorm, ReLU	128	1 × 1 × 512	1 × 1, [0,0,0,0]	28 × 28 × 128
	Conv 2_2_2	Conv, BatchNorm, ReLU	128	3 × 3 × 128	1 × 1, [1,1,1,1]	28 × 28 × 128
	Conv 2_2_3	Conv, BatchNorm	512	1 × 1 × 128	1 × 1, [0,0,0,0]	28 × 28 × 512
Stack 2_2_br	-	-	-	-	-
Stack 2_2_add		Add, ReLU	-	-	-
Stack 2_3	Conv 2_3_1	Conv, BatchNorm, ReLU	128	1 × 1 × 512	1 × 1, [0,0,0,0]	28 × 28 × 128
	Conv 2_3_2	Conv, BatchNorm, ReLU	128	3 × 3 × 128	1 × 1, [1,1,1,1]	28 × 28 × 128
	Conv 2_3_3	Conv, BatchNorm	512	1 × 1 × 128	1 × 1, [0,0,0,0]	28 × 28 × 512
Stack 2_3_br	-	-	-	-	-
Stack 2_3_add		Add, ReLU	-	-	-
Stack 2_4	Conv 2_4_1	Conv, BatchNorm, ReLU	128	1 × 1 × 512	1 × 1, [0,0,0,0]	28 × 28 × 128
	Conv 2_4_2	Conv, BatchNorm, ReLU	128	3 × 3 × 128	1 × 1, [1,1,1,1]	28 × 28 × 128
	Conv 2_4_3	Conv, BatchNorm	512	1 × 1 × 128	1 × 1, [0,0,0,0]	28 × 28 × 512
Stack 2_4_br	-	-	-	-	-
Stack 2_4_add		Add, ReLU	-	-	-

,

Table 3. CNN feedforward calculation in stack 3.

Layer		Operation	Filters’ Set Num	Filter Size	Stride, Zero-Padding [Top, Bottom, Left, Right]	Feature Map Output
Stack 3_1	Conv 3_1_1	Conv, BatchNorm, ReLU	256	1 × 1 × 512	2 × 2, [0,0,0,0]	14 × 14 × 256
	Conv 3_1_2	Conv, BatchNorm, ReLU	256	3 × 3 × 256	1 × 1, [1,1,1,1]	14 × 14 × 256
	Conv 3_1_3	Conv, BatchNorm	1024	1 × 1 × 256	1 × 1, [0,0,0,0]	14 × 14 × 1024
Stack 3_1_br	Conv 3_1_4	Conv, BatchNorm	1024	1 × 1 × 512	2 × 2, [0,0,0,0]
Stack 3_1_add		Add, ReLU	-	-	-
Stack 3_2	Conv 3_2_1	Conv, BatchNorm, ReLU	256	1 × 1 × 1024	1 × 1, [0,0,0,0]	14 × 14 × 256
	Conv 3_2_2	Conv, BatchNorm, ReLU	256	3 × 3 × 256	1 × 1, [1,1,1,1]	14 × 14 × 256
	Conv 3_2_3	Conv, BatchNorm	1024	1 × 1 × 256	1 × 1, [0,0,0,0]	14 × 14 × 1024
Stack 3_2_br	-	-	-	-	-
Stack 3_2_add		Add, ReLU	-	-	-
Stack 3_3	Conv 3_3_1	Conv, BatchNorm, ReLU	256	1 × 1 × 1024	1 × 1, [0,0,0,0]	14 × 14 × 256
	Conv 3_3_2	Conv, BatchNorm, ReLU	256	3 × 3 × 256	1 × 1, [1,1,1,1]	14 × 14 × 256
	Conv 3_3_3	Conv, BatchNorm	1024	1 × 1 × 256	1 × 1, [0,0,0,0]	14 × 14 × 1024
Stack 3_3_br	-	-	-	-	-
Stack 3_3_add		Add, ReLU	-	-	-
Stack 3_4	Conv 3_4_1	Conv, BatchNorm, ReLU	256	1 × 1 × 1024	1 × 1, [0,0,0,0]	14 × 14 × 256
	Conv 3_4_2	Conv, BatchNorm, ReLU	256	3 × 3 × 256	1 × 1, [1,1,1,1]	14 × 14 × 256
	Conv 3_4_3	Conv, BatchNorm	1024	1 × 1 × 256	1 × 1, [0,0,0,0]	14 × 14 × 1024
Stack 3_4_br	-	-	-	-	-
Stack 3_4_add		Add, ReLU	-	-	-
Stack 3_5	Conv 3_5_1	Conv, BatchNorm, ReLU	256	1 × 1 × 1024	1 × 1, [0,0,0,0]	14 × 14 × 256
	Conv 3_5_2	Conv, BatchNorm, ReLU	256	3 × 3 × 256	1 × 1, [1,1,1,1]	14 × 14 × 256
	Conv 3_5_3	Conv, BatchNorm	1024	1 × 1 × 256	1 × 1, [0,0,0,0]	14 × 14 × 1024
Stack 3_5_br	-	-	-	-	-
Stack 3_5_add		Add, ReLU	-	-	-
Stack 3_6	Conv 3_6_1	Conv, BatchNorm, ReLU	256	1 × 1 × 1024	1 × 1, [0,0,0,0]	14 × 14 × 256
	Conv 3_6_2	Conv, BatchNorm, ReLU	256	3 × 3 × 256	1 × 1, [1,1,1,1]	14 × 14 × 256
	Conv 3_6_3	Conv, BatchNorm	1024	1 × 1 × 256	1 × 1, [0,0,0,0]	14 × 14 × 1024
Stack 3_6_br	-	-	-	-	-
Stack 3_6_add		Add, ReLU	-	-	-

and

Table 4. CNN feedforward calculation from stack 4 to output.

Layer		Operation	Filters’ Set Num	Filter Size	Stride, Zero-Padding [Top, Bottom, Left, Right]	Feature Map Output
Stack 4_1	Conv 4_1_1	Conv, BatchNorm, ReLU	512	1 × 1 × 1024	2 × 2, [0,0,0,0]	7 × 7 × 512
	Conv 4_1_2	Conv, BatchNorm, ReLU	512	3 × 3 × 512	1 × 1, [1,1,1,1]	7 × 7 × 512
	Conv 4_1_3	Conv, BatchNorm	2048	1 × 1 × 512	1 × 1, [0,0,0,0]	7 × 7 × 2048
Stack 4_1_br	Conv 4_1_4	Conv, BatchNorm	2048	1 × 1 × 1024	2 × 2, [0,0,0,0]
Stack 4_1_add		Add, ReLU	-	-	-
Stack 4_2	Conv 4_2_1	Conv, BatchNorm, ReLU	512	1 × 1 × 2048	1 × 1, [0,0,0,0]	7 × 7 × 512
	Conv 4_2_2	Conv, BatchNorm, ReLU	512	3 × 3 × 512	1 × 1, [1,1,1,1]	7 × 7 × 512
	Conv 4_2_3	Conv, BatchNorm	2048	1 × 1 × 512	1 × 1, [0,0,0,0]	7 × 7 × 2048
Stack 4_2_br	-	-	-	-	-
Stack 4_2_add		Add, ReLU	-	-	-
Stack 4_3	Conv 4_3_1	Conv, BatchNorm, ReLU	512	1 × 1 × 2048	1 × 1, [0,0,0,0]	7 × 7 × 512
	Conv 4_3_2	Conv, BatchNorm, ReLU	512	3 × 3 × 512	1 × 1, [1,1,1,1]	7 × 7 × 512
	Conv 4_3_3	Conv, BatchNorm	2048	1 × 1 × 512	1 × 1, [0,0,0,0]	7 × 7 × 2048
Stack 4_3_br	-	-	-	-	-
Stack 4_3_add		Add, ReLU	-	-	-
Global average Pool		Global average pooling, Affine	-	7 × 7	-	2048
Classification output		Affine, Soft-Max	-	-	-	3

list the details of the feedforward calculation of the CNN. Figure 8 shows the structural diagram of the CNN corresponding to Table 1, Table 2, Table 3 and Table 4. The CNN consisted of four stacks, as shown in Table 1, Table 2, Table 3 and Table 4 and Figure 8. In stack 1, the channel depth of the feature map increased from 64 to 256. Furthermore, in stack 2, the channel depth increased from 128 to 512. In contrast, for both stacks 1 to 2, the height and width were halved from 56 × 56 to 28 × 28. Similarly, in stacks 3 and 4, the number of channels increased; instead, the height and width of the feature map were downsized.

As shown in Figure 8, the skip structure prevents zero-loss gradients during backpropagation in the training stage, even when the layers are deepened. BatchNorm in Table 1, Table 2, Table 3 and Table 4 refers to the batch normalization operation [19]; this operation allows for accelerated learning and prevents overfitting as a side effect. The BatchNorm operation has been employed in CNN models since the introduction of ResNet [16] and DenseNet [20]. A rectified linear unit (ReLU) function [21] is used in deep neural networks to prevent gradient loss. The ReLU function has been applied in various types of CNNs, such as AlexNet [22] and VGG [23]. The global average pooling operation [24] in Table 4 generates a one-dimensional vector with elements of the channel number. Each element represents the average value of each channel in the feature map. This downsizing of the feature map shortens the computational load and has been used since GoogLeNet [25].

3. Results

The CNN was developed using a MathWorks’ MATLAB (R2022a) and a Deep Learning Toolbox. The CNN was validated with 64 images; i.e., 20 of class A (Good), 20 of B (Neutral), and 24 of C (Bad) images. In machine learning, a sufficient dataset (such as medical images) is not always available. Therefore, various evaluation methods based on limited data have been proposed [26]. In this study, we adopted a repeated random subsampling validation. As shown in Figure 9, three images from each class were randomly extracted to test the CNN after training. Accordingly, the remaining 55 images, excluding nine test images, were used to train the CNN; subsequently, nine test images were input to the trained CNN to confirm the classification accuracy.

The above validation was conducted over 20 trials, as explained in the following section.

3.1. Training of CNN

Table 5. CNN training conditions.

	First Half	Last Half
Solver	SGDM (Stochastic Gradient Descent with Momentum)
Learning Rate	${10}^{-3}$
Total Epochs/Iterations	200	30
Mini batch Size	256	128
Augmentation	Random Left–Right Reflection (50%), Random Top–Down Reflection (50%)
	Random rotation ranged in [0,360] degree	No rotation
CPU	Intel core i9 12900K
Main Memory	128 GB
OS	Windows 11 Pro 64bit
Development Language	MathWorks, MATLAB (R2022a)
GPU	Nvidia RTX A6000 (VRAM 48 GB, 10752 CUDA cores)

lists the conditions for training the CNN. The training was divided into two phases: First Half and Last Half. The training images were augmented, as shown in Figure 10. In the First Half, the image was randomly flipped from left to right (Figure 10a) and then up and down (Figure 10b) and randomly rotated (Figure 10c) in the range of real numbers from 0° to 360°. Random rotation allows training images to be virtually a significant amount of image data [27]. In the Last Half, training is performed on images without rotation to acclimate the CNN to the non-rotated images. In this study, the CNN could detect crack features with many augmented images. Therefore, transfer learning [28], such as fine-tuning a pre-trained CNN, should be the focus of future work.

Figure 11 shows the transition in loss; the cross-entropy function was adopted [29]. In all trials, the loss declined. In particular, in the Last Half (Figure 11b), the loss decrease was smooth because the augmentation was only inverted vertically and horizontally.

3.2. CNN Evaluation

Table 6. Test results of CNN.

Trial	A (Good) IDs for Test			B (Neutral) IDs for Test			C (Bad) IDs for Test			Accuracy
1	3	10	16	1	13	15	7	11	21	1.0000
2	5	8	10	8	18	19	14	16	20	0.8889
3	5	6	11	7	17	18	2	18	21	0.8889
4	2	5	8	5	6	14	3	16	18	0.8889
5	1	16	17	18	19	20	1	5	10	0.8889
6	3	5	19	2	8	18	7	12	19	1.0000
7	9	14	17	3	7	11	11	12	18	0.6667
8	5	10	14	1	2	11	1	8	21	1.0000
9	2	5	10	18	19	20	7	10	22	0.7778
10	5	8	11	2	6	14	7	13	21	1.0000
11	10	16	19	9	12	14	2	13	22	1.0000
12	3	12	15	12	17	19	11	16	22	0.7778
13	7	9	18	1	6	7	6	7	8	0.7778
14	1	9	20	2	11	17	9	13	15	0.7778
15	1	6	14	3	16	17	1	2	24	0.5556
16	2	9	13	2	17	19	3	6	15	0.8889
17	4	8	18	4	8	10	12	23	24	0.8889
18	5	15	18	3	15	20	13	19	21	0.6667
19	13	14	17	1	5	8	1	15	21	1.0000
20	7	10	19	4	13	15	4	15	22	1.0000
–	–			–			–			Average: 0.8667

lists the test data and classification accuracies for each trial. The mean accuracy was 86.67%. The accuracy is plotted in Figure 12a, and the confusion matrix is shown in Figure 12b. As shown in Figure 12b, misjudgment cases between classes A (Good) and C (Bad) were not found.

Table 7. Misclassified image IDs, estimated classes, and probabilities of softmax layer.

			Probability
Image ID	Trial	Misjudged Class	A	B	C
A001	14	B	0.467	0.532	0.001
A001	15	B	0.211	0.787	0.002
A018	13	B	0.017	0.980	0.003
A018	17	B	0.053	0.942	0.004
A018	18	B	0.205	0.793	0.002
B003	7	A	0.911	0.089	0.000
B003	15	A	0.899	0.101	0.000
B003	18	A	0.816	0.184	0.000
B011	14	A	0.589	0.411	0.001
B016	15	C	0.000	0.056	0.944
B017	12	C	0.000	0.465	0.535
B017	15	C	0.000	0.440	0.560
B019	2	C	0.000	0.076	0.924
B019	5	C	0.000	0.296	0.704
B019	9	C	0.000	0.070	0.930
B019	12	C	0.000	0.246	0.754
B019	16	C	0.000	0.233	0.767
B020	9	C	0.000	0.478	0.522
C008	13	B	0.000	0.512	0.488
C012	7	B	0.000	0.559	0.441
C018	3	B	0.002	0.888	0.110
C018	4	B	0.009	0.968	0.023
C018	7	B	0.001	0.973	0.026
C019	18	B	0.000	0.508	0.492

lists the IDs of the misclassified images and the probability values of the softmax layer. Figure 13 shows the misclassified images corresponding to the image IDs listed in Table 7.

The CNN classified B017, B019, and B020 as C (Bad) because the cracks were deep and thick. Such discrimination is difficult even for humans and is attributed to fuzzy boundaries in the classification process [30]. In addition, the probability values of B (Neutral) and C (Bad) in B017, B020, C008, C012, and C019 were simultaneously close to 0.5, indicating that the misjudgment was made in confusion. The CNN was confused and acted like humans in these cases.

It is noteworthy that for A001, where the CNN determined A (Good) to be B (Neutral), we assumed that the annual rings were misrecognized as cracks. This problem can be addressed in the future by adding training image data.

Considering the loss transition during training shown in Figure 11, training was performed successfully. As shown in Table 6 and Figure 12, the CNN could assess the overall crack severity appropriately, even though some misjudged cases were found. Furthermore, considering the results in Table 7 and Figure 13, we conclude that the CNN acts like a human and, therefore, has the potential to evaluate timber cracks on behalf of a human.

We also evaluated the Recall, Precision, and F-measure [31], commonly used to evaluate the classifier for three or more categories. The Recall, Precision, and F-measure are defined in Equations (1)–(3).

Table 8. Results of Recall, Precision, F-measure.

Class	Recall	Precision	F-Measure
A (Good)	$\left(\frac{55}{55+5}\right)=0.917$	$\left(\frac{55}{55+4}\right)=0.932$	0.924
B (Neutral)	$\left(\frac{47}{4+47+9}\right)=0.783$	$\left(\frac{47}{5+47+6}\right)=0.810$	0.797
C (Bad)	$\left(\frac{54}{6+54}\right)=0.900$	$\left(\frac{54}{9+54}\right)=0.857$	0.878

shows the result.

$$\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

(1)

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

(2)

$$\mathrm{F}-\mathrm{measure}=\frac{2 \times \mathrm{Precision} \times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$

(3)

Given the results, we can observe that there is no excessive bias in the data set.

4. Discussion

This paper described the automatic evaluation system of internal cracks in square timbers that occur due to high-temperature drying, which is not externally visible. Currently, the evaluation of internal cracks is performed visually by humans. However, AI could take the place of this process. In this paper, we applied a CNN, widely used in machine learning, for image recognition. In many cases, CNNs are employed for crack detection in wood [13,14,15] and concrete [18] construction materials. Most of the relevant studies in crack recognition using a CNN do not focus on the evaluation of the level of severity of the detected cracks. The present paper differs from other similar studies in that it focuses on evaluating the material quality considering cracks using a CNN. In addition, a CNN could potentially be able to perform crack recognition regardless of light conditions if we increase the training images. Such a task is difficult to carry out using only classical image processing such as edge detection.

Section 2 describes the details of the high-temperature-dried square timber and the proposed CNN. In Section 3, the performance of the CNN’s evaluation ability is validated. The results confirm that the evaluation was similar to that of a human.

To improve the durability of wooden buildings, it is essential to evaluate the quality of their materials, as well as to contribute to sustainability, aesthetics, the environment, and livability. In short, addressing the durability of wooden constructions may encourage architects, urban planners, and policy planners to rethink and reshape streets [32] with sustainable materials. In addition, it could improve the quality, aesthetics, and functional ambiance of urban landscape environments.

5. Conclusions

This paper describes the crack severity level classification of square timber cross-sectional images using a CNN. Thirty-two cedar timbers of 3000 mm in length were dried at a high temperature. Cross-sectional images of two different parts of each timber were obtained. Consequently, 64 cross-sectional images were acquired in total. The images were classified into three classes, i.e., A (Good), B (Neutral), and C (Bad), based on the severity of the cracks. We subsequently investigated the automatic classification of images using the ResNet50-CNN. Because the image data were insufficient, flipping and rotation were applied while training the CNN to virtually augment the image data.

In addition, repeated random subsampling validation was performed to evaluate the CNN using a small amount of data. Three images from each class were randomly selected for testing, and the remaining images were used to train the CNN. The loss decreased as the epochs progressed. The classification accuracy was evaluated using test data that were not used for the training; the aforementioned procedure was performed 20 times. In the 20 trials, the average classification accuracy for the test data was 86.67%. We analyzed the misjudged images and observed confusion similar to human evaluators. Therefore, we confirmed that the CNN has the potential to evaluate cross-sectional cracks on behalf of humans. However, the light condition to capture images was constant in this paper. In the future, it will be necessary to prepare more image data to evaluate the performance of the CNN. In addition, we will need to construct a system in the field that enables the evaluation of cracks under various light conditions.

This study could decrease the burden of quality control on vendors, as well as researchers improving high-temperature drying techniques, who have to rank many square timbers visually. The practical application of this study is to replace humans with artificial intelligence that automatically evaluates the quality of timbers. A skilled human evaluator could comprehensively evaluate timber quality, taking into account not only the cracks number, area, and length, but additionally the shape and location of the cracks in the cross-section. CNNs have the potential to perform the detailed judgment of a skilled evaluator if trained with various image data.