Introduction

During their growth, trees are inevitably affected by internal and external factors, leading to various wood defects, such as knots, cracks, mold, and insect eyes.These defects affect not only the appearance and quality of the wood but also its mechanical properties, including strength and texture, and must be removed during processing. Plywood, one of China’s three major man-made panels, is highly sought after in the market and mainly made of a veneer and adhesives.The quality of China’s plywood production varies, primarily determined by the quality of the veneer, which is influenced by defects such as knots, decay, cracks, insect pests, and scars.Knots and cracks are the most common defects affecting veneer quality. Currently, most production enterprises rely on workers to visually detect. However, manual detection is labor intensive, prone to eye fatigue, and results in low accuracy, low efficiency and high costs.Defects negatively affect the mechanical properties, appearance, quality and utilization of wood. Manual detection increases the production costs of enterprises, making identifying veneer defects via some other means particularly important1. Using computer vision instead of manual inspection can effectively reduces the production costs of enterprises and improves the detection accuracy and detection efficiency of veneer defects2,3. However, due to the collection equipment or factory environments, collected veneer images often have fuzzy edges, inconspicuous contrast and distortion, etc.The direct use of the original veneer images for defect detection can substantially affect the defect detection rates. Therefore, effectively enhancing the original veneer defect image is necessary to minimize environment and equipment interference, ensuring that defective parts and details are more clearly observed.

The rapid development of the computer technology and computer hardware, such as graphics processing units has advanced image enhancement algorithms. Chen et al.4 avoided the problem of excessively large or small thresholds by improving the wavelet threshold function, enhancing the adaptability of the algorithm, and achieving an excellent denoising effect. The traditional histogram equalization (HE) algorithm is not ideal for low-quality image processing.Converting an RGB image into HSV color space and performing weighted histogram equalization cam effectively avoid issues of detail loss and gray scale merging5. Zh et al.6 used wavelet transform to decompose a defect image, applied an improved Retinex algorithm to denoise the low frequencies, used a fuzzy rule to enhance the high frequencies, and performed inverse transform to obtain an enhanced image, achieving a defogging effect. Huo et al.7 first used the CLAHE algorithm to enhance a low-contrast image before splicing it, which effectively increased the number of matching points and improved the quality of the spliced image. Xu et al.8 employed an improved weighted bootstrap filtering algorithm by superimposing an original image with the gradient map, enhancing the contour using the Canny operator, and enhancing image details based on the fuzzy theory, they obtained an enhanced image through BM3D denoising, effectively improving image details and contrast.Therefore ZH et al.9 used a rough set and particle swarm optimization algorithms for the smooth regions, edges, and texture regions of a defect image.The enhanced image showed clearer edges, richer texture details, and retained the information of the smooth region of the original image. Wang et al.10 determined the optimal segmentation threshold of image brightness using a genetic algorithm, divided the brightness channel into multiple image sequences with different exposure levels, applied multi-threshold chunking enhancement to enhance the sub-map, and obtained an enhanced image after multi-scale fusion, effectively restoring the real brightness distribution of the shooting environment. Many scholars have proposed excellent methods tailored to the characteristics of wood science and have applied high-performance algorithms from other fields to wood science, achieving good results. Mou11 used grayscale transformation to enhance a defect image before feature extraction, confirming that the extracted feature parameters were beneficial for subsequent identification tasks.Yang et al.12 used NSST to preprocess wood surface defects, reducing the complexity of defect images and the extent of computation performed.They then extracted the defect features using an improved CNN, used extreme learning machine to classify the defects, and employed a genetic algorithm to optimize the classifier parameters, obtaining high recognition accuracy.Wang et al.13 for the shortcomings of the traditional wavelet threshold denoising algorithm, by improving the Lip index threshold and constructing a new wavelet threshold function, the enhanced image can better retain the detail information, eliminate more noise coefficients, and be closer to the real value.

Although the above methods have some effect on their objects, by the characteristics of the research images in this paper.Due to the influence of environment, lighting, and noise on single board defect images when acquiring them, issues such as low brightness, low contrast, or mixed noise may occur in the collected images. To eliminate irrelevant information in the images, restore useful information, and improve their analyzability and usability, the AMEF algorithm is used to enhance the details of the defect images, achieving the best contrast and brightness, This approach is applicable to all single board defect images.Therefore, the AMEF-AGC algorithms is used for image enhancement processing.To address the issues of blurred edges, low contrast and distortion in veneer defect images, this study combines the AMEF and AGC algorithms to enhance the low-resolution veneer defect images, making the defective parts and details of the image more clear.First, the AMEF algorithm is used to enhance the details of defect images, and the improved AGC algorithm is used to enhance the fused veneer defect images to achieve the best contrast and brightness. The improved AGC algorithm can be used for all veneer defect images without the need for image preclassification effect of the proposed method on defect images is considerably high, thus improving recognition accuracy in subsequent tasks.

Materials and methods

Image acquisition

In this study, the dataset of veneer defects is established by manual acquisition.To simulate the influence of different environments and light, the veneer defect images selected in this project were collected from the laboratory and a furniture production workshop in Hebei at different times, and the wood species were chosen as experimental samples, including rotary cut veneer of Pine, poplar and birch. Due to the influence of environment, lighting, and noise during the acquisition process of single board defect images, problems such as low brightness, low contrast, or mixed noise may occur in the collected images. An example of a defect image is shown in (Fig. 1).

Fig. 1
figure 1

Example of defect images.

Full size image

Establishment of dateset

The acquisition of veneer defect images involves the use of a camera, tripod and computer. The role of each device is as follows: the camera(Canon DS126431) acquires veneer defect images; the resolution of images is 3456 × 4608; the tripod is used to position the camera appropriately; and the computer handles the reception and processing of the required veneer defect images.In this experiment, 450 veneer defect images are collected, among which 150 images of live knots, dead knots and cracks are collected. In order to build a proper experimental dataset, the original images are first cropped to cut off the extra background part of the images. Considering that too large images will increase the complexity of network computing and occupy more memory, all images were cut into image blocks with the size of 128 × 128 pixels. Including dead knots, live knots, and cracks as three kinds of defect images. The three kinds of defects of veneer are schematically shown in (Fig. 2).

Fig. 2
figure 2

Veneer defect image.

Full size image

Table 1 Shows the hardware and software environment configuration adopted in this paper.

Table 1 Hardware environment configuration table.
Full size table

Gamma corrected image enhancement algorithm

Gamma correction (GC) is to adjust the Gamma curve by debugging the γ parameter to change the contrast of the images, which is used to enhance the detail information of the image, and the Gamma Correction is generally accomplished by manually adjusting the value of γ. The GC algorithm is shown in Eq. (1).

$$\:{\text{I}}_{\text{o}\text{u}\text{t}}={\left({\text{I}}_{\text{i}\text{n}}\right)}^{{\upgamma\:}}$$
(1)

Where \(\:{I}_{in}\) and \(\:{I}_{out}\)are the input and output image intensities, and γ is a control parameter in the transformation process, it plays a key role in the algorithm, and this parameter determines the final enhancement effect.

The effect of Gamma correction is shown in Fig. 3, with γ values of 0.3, 0.6, 1, 1.5 and 2.

Fig. 3
figure 3

Comparison of the image enhancement effect by the parameter value.

Full size image

Whenγ>1, the overall brightness of the image decreases, and the visual effect becomes dark; whenγ<1, the overall brightness of the image is enhanced, and the visual effect becomes bright.

Adaptive gamma correction image enhancement algorithm

The gamma correction enhancement algorithm relies heavily on parameter and a single approach for image processing, which can easily lead to the over-enhancement of bright areas and white visual effects.Furthermore, manually adjust parameters is impractical when the number of images is large. When the AGC algorithm was proposed, the aforementioned problems were addressed.Thus, the AGC algorithm is efficient, convenient, and stable as well as widely studied. The AGC enhancement algorithm is shown in Eq. (2).

$$\:{\text{I}}_{\text{o}\text{u}\text{t}}=\:\text{c}\:{\left({\text{I}}_{\text{i}\text{n}}\right)}^{{\upgamma\:}}$$
(2)

Compared with Eq. (1), the parameter c is added, and the parameter c is used to modify the image lightness and darkness.

In experiments, we found that different enhancement processes should be performed depending on the brightness of the image. Therefore, the veneer defect images are categorized into high contrast high brightness images(HCHB), high contrast low brightness images(HCLB), low contrast high brightness images(LCHB), and low contrast low brightness images(LCLB).We selected 12 defective images from the established database with three different defect types for each category and Figs. 4, 5, 6 and 7 show the image enhancement results for LCLB, LCHB, HCLB, and HCHB.

Fig. 4
figure 4

Before and after contrast of three types of single plate defects with low contrast and low brightness.

Full size image

The effect of the AGC algorithm is shown in 4b, 4d and 4 f. The visual effect of the enhanced image is obviously improved, and the algorithm has good performance for these kind of images.Therefore, the veneer defect images are categorized into high contrast high brightness.

Fig. 5
figure 5

Before and after contrast of three types of single plate defects with low contrast and high brightness.

Full size image

The effect of the AGC algorithm is shown in 5b, 5d and 5 f. It is obvious that the AGC processing obviously does not achieve satisfactory results, and the processed image becomes darker.

Fig. 6
figure 6

The contrast between three types of single plate defects with high contrast and low brightness.

Full size image

The effect of the AGC algorithm is shown in 6b, 6d and 6f, and the human eye can observe that the quality of the image has been significantly improved, and more detailed information can be seen after enhancement.

Fig. 7
figure 7

The contrast between three types of single plate defects with high contrast and high brightness.

Full size image

The effect of the AGC algorithm is shown in 7b, 7d and 7f, and AGC does not have a significant enhancement effect on the image of HCHB, with low robustness.

In summary, while the AGC algorithm offers several advantages, it has some drawbacks.The enhancement of veneer defect images is poor, particularly when the image are near the demarcation line or they are LCHB image, where the quality may not improve and could even deteriorate. This limits the application prospects of the algorithm.Therefore, improvements to the algorithm are proposed herein to address these drawbacks.

Improved AGC of image enhancement algorithm

The AGC algorithm is introduced in Sect. 2.4.Although it is efficient, the enhancement results are poor when the mean or standard deviation values of defect images are close to the separation boundaries for each category. This is because the enhancement function changes abruptly from low contrast to high contrast or from low brightness to high brightness. To address this challenge, this section proposes to compute γ and c based on nonlinear weights and modify the enhancement function to be continuous to process images located on the boundary of two classes and better enhance them.

Constructing a nonlinear weight adjustment function, the Construct a discontinuous function f (x), S(x) is a step function, f (x) = \(\:{\text{f}}_{1}\)(x) when x ≤ \(\:{\text{x}}_{0}\), and f (x) = f2(x) when x > \(\:{\text{x}}_{0}\). f (x) can be compared with S(x).when x≤\(\:{\text{x}}_{0}\), f (x)=f2(x). f (x) and S(x) can be written as Eq. (3) and Eq. (4).

$$\:\text{S}\left(\text{x}\right)=\left\{\begin{array}{cc}0,&\:\text{x}\le\:{\text{x}}_{0}\\\:1,&\:\text{x}>{\text{x}}_{0}\end{array}\right.$$
(3)
$$\:\text{f}\left(\text{x}\right)=\left(\begin{array}{c}1-\text{S}\left(\text{x}\right)\end{array}\right){\text{f}}_{1}\left(\text{x}\right)+\text{S}\left(\text{x}\right){\text{f}}_{2}\left(\text{x}\right)$$
(4)

S(x) can change the properties of f(x).To make f(x) a continuous function, a continuous nonlinear function is constructed as shown in Eq. (5). Figure 8 shows the two forms of the S(x) function for \(\:{\text{x}}_{0}\)= 0.3. Equation (6) can change the steepness of the function by adding the parameter p to the exponential part. As p increases, the function image becomes steeper and approaches a step function.

$$\:\text{S}\left(\text{x}\right)={\text{x}}^{-\text{l}\text{o}{\text{g}}_{2}\left(\text{x}\right)}$$
(5)
$$\:\text{S}\left(\text{x}\right)={\text{x}}^{-\text{p}\text{l}\text{o}{\text{g}}_{2}\left(\text{x}\right)}$$
(6)
Fig. 8
figure 8

Step function and the near ladder function are represented graphically.

Full size image

Next, the enhancement parameters are improved, based on the construction of the nonlinear weight tuning function, and similarly, the nonlinear weight tuning function \(\:{w}_{\gamma\:}\)for the contrast enhancement function γ as well as the continuous contrast enhancement function\(\:{\:\:\gamma\:}_{v}\)can be obtained.The formulas are given in (7) and (8).

$$\:{\text{w}}_{{\upgamma\:}}={\left(2{\upsigma\:}\right)}^{-{\text{p}}_{{\upgamma\:}}\text{l}\text{o}{\text{g}}_{2}\left(2{\upsigma\:}\right)}$$
(7)
$$\:{\:\:{\upgamma\:}}_{\text{v}}={\text{w}}_{{\upgamma\:}}{\text{e}}^{\frac{1-\left({\upmu\:}+{\upsigma\:}\right)}{2}}+\left(1-{\text{w}}_{{\upgamma\:}}\right)\left(-{\text{l}\text{o}\text{g}}_{2}{\upsigma\:}\right)$$
(8)

As \(\:\sigma\:\:\)increases, \(\:{w}_{\gamma\:}\)also increases, and when \(\:\sigma\:\)=0.5 \(\:{w}_{\gamma\:}\) = 1, at which point Eq. (8) converges as shown in (9) and (10).

$$\:\underset{{\upsigma\:}\to\:0}{\text{l}\text{i}\text{m}}\:{{\upgamma\:}}_{\text{v}}=-{\text{l}\text{o}\text{g}}_{2}\:\sigma\:$$
(9)
$$\:\underset{{\upsigma\:}\to\:0.5}{\text{l}\text{i}\text{m}}\:{{\upgamma\:}}_{\text{v}}={\text{e}}^{\frac{1-\left({\upmu\:}+{\upsigma\:}\right)}{2}}$$
(10)

Similarly, the nonlinear weight adjustment function \(\:{\text{w}}_{\text{c}}\)for the luminance adjustment function c and the continuous luminance adjustment function \(\:{\text{c}}_{\text{v}}\) are obtained. The formulas are given in (11)–(13).

$$\:{\text{w}}_{\text{c}}={{\upmu\:}}^{-{\text{p}}_{\text{c}}{\text{l}\text{o}\text{g}}_{2}{\upmu\:}}$$
(11)
$$\:{\text{c}}_{\text{v}}=\frac{1}{{\text{w}}_{\text{c}}+\left(1-{\text{w}}_{\text{c}}\right){\text{k}}_{\text{v}}}$$
(12)
$$\:{\text{k}}_{\text{v}}={\text{I}}_{\text{i}\text{n}}^{{{\upgamma\:}}_{\text{v}}}+\left(1-{\text{I}}_{\text{i}\text{n}}^{{{\upgamma\:}}_{\text{v}}}\right){{\upmu\:}}^{{{\upgamma\:}}_{\text{v}}}$$
(13)

As µ decreases, the weight of \(\:{w}_{c}\)decreases and (1-\(\:{w}_{c}\))\(\:{k}_{v}\)dominates in Eq. (12).As µ increases, the weight of \(\:{w}_{c}\) increases, and when µ = 1, \(\:{w}_{c}\) = 1, \(\:{c}_{v}\) = 1. This is shown in Eqs. (14) and (15).

$$\:\underset{{\upmu\:}\to\:0}{\text{l}\text{i}\text{m}}{\text{c}}_{\text{v}}=\frac{1}{{\text{k}}_{\text{v}}}=\frac{1}{{\text{I}}_{\text{i}\text{n}}^{{{\upgamma\:}}_{\text{v}}}+\left(1-{\text{I}}_{\text{i}\text{n}}^{{{\upgamma\:}}_{\text{v}}}\right){{\upmu\:}}^{{{\upgamma\:}}_{\text{v}}}}$$
(14)
$$\:\underset{{\upmu\:}\to\:1}{\text{l}\text{i}\text{m}}{\text{c}}_{\text{v}}=\frac{1}{{\text{w}}_{\text{c}}}=1$$
(15)

Finally, the image augmentation is achieved by bringing the nonlinear weight-adjusted \(\:{\:\:\gamma\:}_{v}\) and from \(\:{c}_{v}\) into Eq. (16).

$$\:{\text{I}}_{\text{o}\text{u}\text{t}}^{{\prime\:}}={\text{c}}_{\text{v}}{\text{I}}_{\text{i}\text{n}}^{{{\upgamma\:}}_{\text{v}}}$$
(16)

Next, the AGC algorithm and the improved AGC algorithm are used to enhance the four types of veneer defect images of LCLB, LCHB, HCLB, and HCHB, and the images before and after the enhancement and the corresponding histograms are shown in Figs. 9, 10 and 11, and Fig. 12 (the improved AGC algorithm is referred to as AGC+).

Fig. 9
figure 9

Comparison of LCLB image AGC with AGC + enhancement effect and histogram.

Full size image
Fig. 10
figure 10

LCHB image AGC compared with AGC + enhancement effect and histogram.

Full size image
Fig. 11
figure 11

Comparison of HCLB image AGC with AGC + enhancement effect and histogra.

Full size image
Fig. 12
figure 12

Comparison of HCHB image AGC with AGC + enhancement effect and histogram.

Full size image

It can be observed that compared with the AGC + algorithm, AGC does not improve the contrast and brightness of the defective images, with preclassified boundaries.The histogram distribution is more uneven and shifted to the lower gray levels, making the images darker and the enhancement result rather inferior compared to the original image.However, the AGC + algorithm notably improves the contrast and brightness of defect images, leads to a more uniform histogram distribution, and allows more information to be obtained from the images.

AMEF image enhancement algorithm

An unavoidable problems arises during image acquisition: factory production generates a large amount of dust and floating impurities in the environment, which causes light scattering.The prevents the image acquisition equipment from achieving the best exposure, reduces image visibility, causes the loss of the image details, and affects overall image quality.The AMEF algorithm is mostly used for image defogging.After experiments, it was found to demonstrate good reduction effect on the loss of image details and visibility reduction caused by the aforementioned reasons.The principle of the AMEF algorithm is shown in Eq. (17).

$$\:\begin{array}{c}J\left(\text{x}\right)=\sum\:_{\text{k}=1}^{\text{K}}\:{\text{W}}_{\text{K}}\left(\text{x}\right){\text{E}}_{\text{K}}\left(\text{x}\right)\end{array}$$
(17)

Where J(x) is the fused image, K represents the number of differently exposed E(x)images, and \(\:{W}_{K}\) represents the pixel weight map.

The specific experimental steps for the AMEF algorithm are as follows:

1.The original veneer defect image is processed with Gamma correction to obtain k underexposed veneer defect images.

2.For each underexposed veneer defect image derive its pixel-level weight map as shown in Eq. (18).

$$\:\begin{array}{c}{\text{W}}_{\text{K}}\left(\text{x}\right)={\text{C}}^{\text{k}}\left(\text{x}\right)\cdot\:{\text{S}}^{\text{k}}\left(\text{x}\right)\end{array}$$
(18)

Previously reported algorithm required block-by-block computation to robustly estimate contrast and saturation, usually through bootstrap filters or similar techniques, which lead to cumbersome computations and unstable performance14. Here, we address these issues by simplifying the computation process.We define the contrast of a pixel at x, \(\:{\text{C}}_{\text{k}}\left(\text{x}\right)\)as a response to a simple Laplace filter, and the saturation of a pixel at x, \(\:{\text{S}}_{\text{k}}\left(\text{x}\right)\)as the standard deviation estimate of the color channel, given the original image, as shown in Eqs. (19) and (20).

$$\:\begin{array}{c}{\text{C}}_{\text{k}}\left(\text{x}\right)=\frac{{\partial\:}^{2}{\text{E}}_{\text{k}}}{\partial\:{\text{x}}^{2}}\left(\text{x}\right)+\frac{{\partial\:}^{2}{\text{E}}_{\text{k}}}{\partial\:{\text{y}}^{2}}\end{array}$$
(19)
$$\:\begin{array}{c}{\text{S}}_{\text{k}}\left(\text{x}\right)=\sum\:_{\text{c}\in\:\text{R},\text{G},\text{B}}\:{\left({\text{E}}_{\text{k}}^{\text{c}}\left(\text{x}\right)-\frac{{\text{E}}_{\text{k}}^{\text{R}}\left(\text{x}\right)+{\text{E}}_{\text{k}}^{\text{G}}\left(\text{x}\right)+{\text{E}}_{\text{k}}^{\text{B}}\left(\text{x}\right)}{3}\right)}^{2}\end{array}$$
(20)

3.In order to spatially extract the detailed features of different resolution images, a Gaussian pyramid is constructed for Wk as shown in Eq. (21).

$$\:\begin{array}{c}{\text{W}}_{\text{k}}^{\text{i}}={\text{d}\text{s}}_{2}\left[{\text{W}}_{\text{k}}^{\text{i}-1}\right]\end{array}$$
(21)

Where \(\:{\text{d}\text{s}}_{2}\)denotes that the image is convolved with a Gaussian kernel and then downsampled to change it to half its original size.

4.Construct a Gaussian pyramid, or Laplace pyramid for the sequence of underexposed images. Equations (22) and (23) is used to produce a set of multi-scale images.

$$\:\begin{array}{c}{\text{E}}_{\text{k}}^{\text{i}}={\text{d}\text{s}}_{2}\left[{\text{E}}_{\text{k}}^{\text{i}-1}\right]\end{array}$$
(22)
$$\:\begin{array}{c}{\text{L}}_{\text{k}}^{\text{i}}={\text{E}}_{\text{k}}^{\text{i}}-{\text{u}\text{s}}_{2}\left[{\text{E}}_{\text{k}}^{\text{i}+1}\right]\end{array}$$
(23)

Where \(\:{us}_{2}\) is the one that upsamples the image to twice its original size.

5.Finally, multiscale fusion is performed to obtain an enhanced defect image, assuming that the size of Ek(x) is m×n. The multiscale combination of all Ek(x)values is achieved by combining each image in the pyramid across the levels using up-sampling summation, as shown in Eq. (24).

$$\:\begin{array}{c}J\left(x\right)={\text{u}\text{s}}_{\left(\text{m},\text{n}\right)}\left[{\text{L}}_{1}^{1}\left(\text{x}\right)\cdot\:{\text{W}}_{1}^{1}\left(\text{x}\right)+\dots\:+{\text{L}}_{\text{K}}^{1}\left(\text{x}\right){\text{W}}_{\text{K}}^{1}\left(\text{x}\right)\right]+\dots\:+\\\:{\text{us}}_{\left(\text{m},\text{n}\right)}\left[{\text{L}}_{1}^{\text{N}}\left(\text{x}\right)\cdot\:{\text{W}}_{1}^{\text{N}}\left(\text{x}\right)+\dots\:+{\text{L}}_{\text{K}}^{\text{N}}\left(\text{x}\right){\text{W}}_{\text{K}}^{\text{N}}\left(\text{x}\right)\right]={\text{us}}_{\left(\text{m},\text{n}\right)}\sum\:_{\text{i}=1}^{\text{N}}\:\left[\sum\:_{\text{k}=1}^{\text{K}}\:{\text{L}}_{\text{k}}^{\text{i}}\left(\text{x}\right)\cdot\:{\text{W}}_{\text{k}}^{\text{i}}\left(\text{x}\right)\right]\end{array}$$
(24)

The enhanced defect image processed using the AMEF algorithm and its local zoomed-in details are shown in Fig. 13. After enhancement using the proposed method, the contour and texture of the enhanced defect image become clearer than those of the original image.Furthermore, the overall quality of the image is improved, making it easier to recognize the defect types.

Fig. 13
figure 13

Comparison Diagram Before and After AMEF Treatment.

Full size image

Based on the above principles, the parameters involved in the application process are provided below. The artificial exposure range selected for Gamma correction is γ∈ {1,2,3,4,5}. Formulas (21), (22), and (23) use conventional separable filters, and the kernel is set to G=[1/16,1/4,3,3,1,4,1,1/16], where the number of layers N in pyramid decomposition is automatically selected based on the image size.

Image enhancement algorithm based on AMEF-AGC+

The AMEF algorithm notably enhances veneer defect images; however, based on the choice of exposure parameters, the enhanced images may be dark. For all parameter selection ranges, a detailed explanation is provided in (Table 2).

Table 2 Explanation of parameter selection.
Full size table

If an input image contains few regions with good contrast, the enhanced image may appear overall darker.Therefore, herein, a defect image enhanced using AMEF was processed using the improved AGC algorithm (AGC+) to optimize image contrast and brightness.

The effect of this paper’s method and the comparison method to enhance the defective image is shown in Figs. 14, 15, 16, and a total of six enhancement algorithms, namely, GC, AGC, improved AGC, AMEF, AMEF and improved AGC, are used for comparison and analysis (for the convenience of labeling, improved AGC algorithm is referred to as AGC+).

Fig. 14
figure 14

Contrast of enhancements to live images.

Full size image
Fig. 15
figure 15

Contrast of enhancements to dead image.

Full size image
Fig. 16
figure 16

Contrast of enhanced effects of crack images.

Full size image

Figure 14–16 show that a defect image enhanced by the improved AGC algorithm exhibits no improvement in detail; Some images processed by GC have excessive contrast enhancement and abnormal color display. The AMEF algorithm has a significant enhancement effect on defect images, but due to the selection of exposure parameters, the enhanced image may appear darker. If the input image contains fewer areas with good contrast, the overall enhanced image may appear darker. Therefore, the proposed method enhances defect images using the AMEF algorithm and then uses a non-linear weight adjustment AGC algorithm for further enhancement.This approach improves image contrast and brightness, enhances defect contour and texture, enriches image detail, and exhibits overall excellent human visual perception.

Results

Evaluation index

This study focuses on the enhancement of veneer defect images. In order to accurately analyze the quality of the enhanced defect images, we used six evaluation indexes: Peak signal-to- noise ratio (PSNR)15,16, structural similarity index measure(SSIM) 17,18, entropy, E19,20, root mean square error (RMSE) [21,22, average gradient (AG)23,24, enhancement measure evaluation (EME) 25. Among them, PSNR, SSIM, and RMSE can reflect the effect of overall image enhancement effect, E reflects image contrast, AG is an indicator of image sharpness, and EME is an indicator of image details. The values are all the larger the better.

PSNR measures the degree of distortion of enhanced images compared to the original images, a high PSNR value indicates lesser distortion between the enhanced and original images, resulting in better image quality, as shown in Eqs. (25) and (26):

$$\:\begin{array}{c}MSE=\frac{1}{K}\sum\:_{k=1}^{K}\:(I\left(i\right)-ˆI\left(i\right){)}^{2}\end{array}$$
(25)
$$\:\begin{array}{c}PSNR=10{\text{log}}_{10}\left(\frac{{MAX}^{2}}{MSE}\right)\end{array}$$
(26)

SSIM reflects the degree of similarity between the original and enhanced images through three aspects: contrast, brightness, and structure.Three aspects to measure the degree of similarity between the original image and the enhanced image.The larger the value of SSIM, the better the quality of the image better, as shown in Equations(27):

$$\:\begin{array}{c}SSIM\left(x,y\right)=\frac{\left(2{\delta\:}_{x}{\delta\:}_{y}+{c}_{1}\right)\left(2{\gamma\:}_{xy}+{c}_{2}\right)}{\left({\delta\:}_{x}^{2}+{\delta\:}_{y}^{2}+{c}_{1}\right)\left({\gamma\:}_{x}^{2}+{\gamma\:}_{y}^{2}+{c}_{2}\right)}\end{array}$$
(27)

Where, x, y are the original and enhanced images, and y are the mean values of the image pixels, and y is the standard deviation of the image pixels, and y is the covariance of x and y, C1 and C2 are parameters.

Entropy is a statistical form of characterization in which the entropy value reflects how much information, on average, is in the image.The greater the uncertainty of the variable, the greater the entropy value, indicating that the image contains more information and has better quality, as shown in Equations(28):

$$\:\text{E}=-\sum\:_{\text{i}=0}^{{2}^{\text{n}}-1}{\text{p}}_{\text{i}}\text{l}\text{o}\text{g}{\text{p}}_{\text{i}}$$
(28)

RMSE is often used as a measure of the contrast of an image.m and n represent the length and width of the image, Iij is the value of the pixel point in row i and column j, µ is the mean value of the image, as shown in Equations(29):

$$\:\text{R}\text{M}\text{S}=\sqrt{\frac{1}{\text{M}\text{N}}\sum\:_{\text{i}=1}^{\text{M}}\sum\:_{\text{j}=1}^{\text{N}}{\left({\upmu\:}-{\text{I}}_{\text{i}\text{j}}\right)}^{2}}$$
(29)

AG is the average value of all pixel points on the gradient map of an image. Average value, which reflects the minute details and texture variations of the image, the sharpness of the image. The larger the average gradient, the image is rich in layers and the clearer the image, as shown in Equations(30):

$$\:\nabla\:\stackrel{⃐}{\text{g}}=\frac{1}{\text{M}\times\:\text{N}}\sum\:_{\text{i}=0}^{\text{M}-1}\sum\:_{\text{j}=0}^{\text{N}-1}\sqrt{\frac{\left({\Delta\:}{\text{I}}_{\text{x}}^{2}+{\Delta\:}{\text{I}}_{\text{y}}^{2}\right)}{2}}$$
(30)

EME is the process of dividing an image into M×N blocks of regions and calculating the logarithmic mean of the ratio of the maximum to the minimum values of the grayscale of all the regions of the map. The larger the EME is, the stronger the local grayscale variation is, and the more detail the image is The larger the EME, the stronger the local grayscale variation, and the richer the details of the image, as shown in Equations(31):

$$\:\text{E}\text{M}{\text{E}}_{\text{N},\text{M}}=\frac{1}{\text{M}\times\:\text{N}}\sum\:_{\text{i}=0}^{\text{M}-1}\sum\:_{\text{j}=0}^{\text{N}-1}20\text{l}\text{o}\text{g}\frac{{\text{I}}_{\text{m}\text{a}\text{x};\text{k},\text{l}}^{{\upomega\:}}}{{\text{I}}_{\text{m}\text{i}\text{n};\text{k},\text{l}}^{{\upomega\:}}}$$
(31)

Experimental results

The quality of the enhanced veneer defect images is evaluated objectively to verify whether the proposed AMEF-AGC + algorithm has a better enhancement effect than those of other algorithms. The specific values are shown in (Table 3).

Table 3 Objective evaluation index results of this thesis method and comparison method.
Full size table

As shown in Table 3, the PSNR, SSIM, E, RMSE, AG, and EME values after enhancement using the proposed method are 49.98, 0.78, 7.22, 42.62, 15.62, and 30.02,respectively.The proposed method achieved the highest values of PSNR, SSIM, AG, and EME.However, the highest scores for the evaluation indices E and RMSE are achieved by the GC algorithm.These results combined with the effect diagrams of the methods employed herein(Figs. 14–16) indicate that images processed by the GC algorithm may experience an excessive enhancement of contrast and brightness, leading to color bias and negative enhancement effects.However, the proposed method offers a more balanced enhancement effect on the defect image, which differs from that of the GC algorithm, and does not produce excessive enhancement.

This study is based on a combination of the AMEF-AGC + algorithm for enhancing veneer defect images.The proposed method is demonstrated to be the best choice for subjective and objective evaluations compared to traditional methods.

Discussion and conclusions

To address various problems pertaining to the acquisition of veneer defect images, using veneer defect images as the research object, the AMEF algorithm was employed to enhance the images, and the AGC algorithm based on nonlinear weight adjustment was used to enhance the contrast and brightness of the images that may be affected by light, noise, or other environmental problems during acquisition. By comparing several algorithms, such as AMEF, AGC, improved AGC, and GC, and evaluating the enhanced defect images using carious evaluation indexes, we verified that the proposed method was the most effective. The proposed method enhanced the fuzzy edges of the veneer defect images, and made other details clearly visible, consequently, the defects were displayed more clearly, and the defect color was more realistic, achieving a good visual effect.Simultaneously, the contrast and brightness of the images were optimally adjusted, improving the quality of the veneer defect images to a certain extent. Improving the usability of defect images lays the foundation for the subsequent accurate recognition of tasks of defect types.

Although results proved the effectiveness of the classification model, obtaining samples pf carious species was challenging.Thus, the training samples used herein was limited to three types of species: poplar, camphor pine, and birch.Future research should aim to improve the algorithm to achieve enhanced performance.