Identification and Classification of Aluminum Scrap Grades Based on the Resnet18 Model

Abstract

In order to reduce the elemental species produced in the recycling and melting of aluminum scrap and to improve the quality of pure aluminum and aluminum alloys, it is necessary to classify the different grades of aluminum scrap before melting. For the problem of classifying different grades of aluminum scrap, most existing studies are conducted using laser-induced breakdown spectroscopy for identification and classification, which requires a clean and flat metal surface and enormous equipment costs. In this study, we propose a new classification and identification method for different grades of aluminum scrap based on the ResNet18 network model, which improves the identification efficiency and reduces the equipment cost. The objects of this research are three grades of aluminum scrap: 1060, 5052, and 6061. The surface features of the three grades were compared using a machine vision algorithm; three different datasets, using RGB, HSV, and LBP, were built for comparison to find the best training dataset for subsequent datasets, and the hyperparameters of learning rate and batch size were tuned for the ResNet18 model. The results show that there was a differentiation threshold between different grades through the comparison of surface features; the ResNet18 network model trained the three datasets, and the results showed that RGB was the best dataset. With hyperparameter optimization of the ResNet18 model, the accuracy of final classification and recognition could reach 100% and effectively achieve the classification of different grades of aluminum scrap.

1. Introduction

Every year, tens of millions of cars in the world reach the end of their service life, which produces tens of millions of tons of automotive scrap; in recent years, countries have been devoted to research on automotive recycling [1,2]. In China, there will be approximately 15 million end-of-life vehicles in 2022, which will generate a large amount of metal scrap. As an essential source of raw materials in the recycling economy, the metal of end-of-life vehicles has the characteristics of large stock, high resource value, and high plasticity.

There are steel, nonferrous metals, precious metals, and other renewable materials in automotive scrap. The separation and sorting of different metals in the scrap are the focus of research at home and abroad. One study [3] used deep learning and transfer learning to identify and separate small samples of nonferrous metals, and the improved YOLOv3 algorithm model was used to identify and classify nonferrous metals in copper and aluminum scrap, with accuracy rates of 91.4% and 95.3%, respectively; another study [4] combined a neural network with automatic sorting of metal scraps, which reduced the data analysis time; to improve the sorting efficiency, researchers proposed scrap aluminum robot sorting technology and evaluated its economics [5].

Most of the current studies on the classification of metal scrap remain on discrimination between different metals and classifying them by their color, shape, and other characteristics, and there are fewer studies on the classification of similar metals. Aluminum is an important nonferrous metal, and the recycling classification of aluminum scrap is one of the research focuses [6]. In this paper, systematic studies were conducted for different grades of aluminum pieces in order to identify higher grade aluminum scrap chips, of which the main components are cast aluminum alloy, alloy aluminum, pure aluminum, etc., the first two of which have numerous grades, and it is still difficult to classify them by grade. In the pretreatment of aluminum scrap, large recycled aluminum plants generally only remove mixed soil and other impurities by sieving, and then directly melt the scrap in the furnace; while for small recycled aluminum plants, this kind of scrap aluminum should be manually divided into foundry aluminum alloy, alloy aluminum, and pure aluminum, and then utilized separately. One study [7] combined deep learning and computer vision for aluminum scrap separation. Some researchers also stratified and separated non-recyclable aluminum scrap using the hydrothermal method [8].

The purpose of advanced aluminum scrap pretreatment technology is to realize the mechanization and automation of aluminum scrap sorting, maximize the removal of metallic and non-metallic impurities, and effectively classify aluminum scrap by alloy composition. The ideal sorting method is to divide aluminum scrap into several categories based on main alloy composition, such as aluminum alloy, aluminum–magnesium alloy, aluminum–copper alloy, aluminum–zinc alloy, aluminum–silicon alloy, etc. This can alleviate the technological difficulty of removing impurities and adjusting the composition in the melting process, and can make comprehensive use of the alloy components in the scrap aluminum, especially scrap aluminum with high zinc, copper, and magnesium content, which need to be stored separately and can be used as intermediate alloy raw materials for melting aluminum alloys.

At present, advanced scrap aluminum sorting technologies mainly include: Eddy Current Separation (ECS) technology [9,10,11,12,13]: by generating a changing magnetic field, which acts on the input scrap; due to the different conductivity and density of different types of aluminum scrap, the aluminum scrap is thrown out at different distances, and sorting can be completed by placing the collection device in advance, but this technology struggles to deal with small amounts of aluminum alloys; X-ray Fluorescence (XRF) analysis: Through the fluorescent radiation generated by emitted X-rays and detected with a solid-state photocathode detector, the concentration of elements in aluminum scrap can be obtained to classify the different alloy components, but it is not currently universally applicable due to the strict X-ray control needed; Laser-Induced Breakdown Spectroscopy (LIBS): With the rapid development of laser-induced breakdown spectroscopy in recent years [14,15,16], the most important present-day research on the classification of aluminum scrap uses laser-induced breakdown spectroscopy, sometimes combined with deep learning, machine learning, and neural networks, which can realize the rapid detection and classification of aluminum scrap but requires a clean and flat aluminum surface and has a high cost. All the above classification techniques have certain defects. The deep learning-based method proposed in this paper effectively avoids these problems. It has a lower operating cost, a faster detection speed, and can realize the classification of large batches, which has more universal applications.

This paper conducted classification experiments on three grades (1060, 5052, 6061) of aluminum blocks: 1060 is pure aluminum, 5052 is alloy aluminum with high magnesium content, and 6061 is alloy aluminum with high magnesium and silicon contents. From the elemental composition, it can be seen that 1060 is very different from 5052 and 6061, while 5052 is not very different from 6061. As the difference between 5052 and 6061 is little, if we can realize the classification of 5052 and 6061, this study has strong practicality. This paper compared different grades of aluminum blocks in terms of light source, surface features, and color space using machine vision algorithms, created different datasets for the above-mentioned different grades of aluminum blocks, trained the different datasets using deep learning algorithms, selected the dataset with the highest correct model training rate as the best dataset for this paper, and used it as the subsequent training dataset. The ResNet18 network model was used to train the dataset, and the accuracy and stability of the model were improved by hyperparameter optimization. The experimental results show that the combination of machine vision and deep learning could effectively classify different grades of aluminum blocks.

2. Data Pre-Processing and Dataset Creation for Aluminum

2.1. Image Acquisition and Preprocessing

Metals are mainly detected and classified by using machine vision to detect the texture features of images of the metal surface [17,18]. The images acquired by the experimental platform used in this paper are shown in Figure 1.

The acquired image was separated from the background, and the target aluminum block was cropped via graying, binarization, finding contours, and other operations; the specific implementation process is shown in Figure 2.

The experimental materials in this paper were aluminum pieces of approximately 20 × 30 mm in size, with grades 1060, 5052, and 6061. Before the experiment, the samples were simply washed with water and dried. Since the aluminum pieces had a silver-white luster, they were placed under the irradiation of the white ring light source shown in Figure 3.

Block images of the three grades of aluminum were acquired separately, and the acquired results are shown in Figure 4.

It can be seen that the aluminum blocks with grades 5052 and 6061 had fewer surface reflections and one can observe surface texture features, while the aluminum block with grade 1060 had the most serious reflections and blurred texture features due to possessing the highest purity of the aluminum. In studying the metal reflection problem, the authors of [19] proposed a new polarization vision system that effectively filtered the reflection noise. In this paper, to solve the reflection problem, considering that the vertical irradiation of light is more concentrated and therefore may lead to more serious reflection, the original vertical irradiation was changed to 45° angle irradiation using a bar-shaped light source. Experiments showed that when the light source is white light, the reflection of the bar shape will still be visible; the authors of [20,21] studied different light sources, and found that a blue light source of wavelength 425~480 was more suitable for the detection of silver background metal; therefore, this experiment used a blue light source at an angle of 45°, as shown in Figure 5, and again collected images of the three grades of aluminum. The results are shown in Figure 6; the visible aluminum block surface reflections disappeared, but the surface clarity was slightly decreased compared to the white ring light source.

2.2. Dataset Image Comparison

LBP (Local Binary Pattern) was the operator used to describe the local texture features of an image with multi-resolution, rotation invariance, gray scale invariance, etc. The main mechanism is to process the surrounding pixels using a threshold and to label the pixels of an image with binary numbers.

In the LBP feature, the joint distribution function T of the local neighborhood is used to represent the texture feature of the image. Assuming that the number of pixels in the local neighborhood is P (P > 1), then T can be expressed as

$$T=t\left(g_c,g_0,\dots ,g_{P-1}\right) P=0, \dots ,P-1$$

(1)

where $g_c$ denotes the local neighborhood center pixel grayscale value and $g_P$ denotes the neighborhood pixel grayscale value.

Assuming that the central pixel and the local neighborhood pixels are independent of each other, T can be expressed as

$$\begin{gathered} T=t\left(g_c,g_0-g_c, \dots ,g_{P-1}-g_c\right) P=0, \dots ,P-1 \\ \approx t\left(g_c\right)\times t\left(g_0-g_c, \dots ,g_{P-1}-g_c\right) \end{gathered}$$

(2)

where $t\left(g_c\right)$ determines the overall brightness of the local area, which is independent of the texture; if it is ignored, T can be obtained as

$$T\approx t\left(g_0-g_c, \dots ,g_{P-1}-g_c\right) P=0, \dots ,P-1$$

(3)

The final result is obtained as

$$T\approx t\left(s\left(g_0-g_c\right), \dots ,s(g_{P-1}-g_c)\right) P=0, \dots P-1$$

(4)

where $s\left(x\right)=\left\{\begin{array}{c}1,x\ge 0\\ 0,x<0\end{array}\right.$, i.e., the defined $LB{P}_{P,R}$, is calculated as

$$LB{P}_{P,R}=\stackrel{P-1}{\underset{P=0}{\sum s(}}{g}_p-g_c)2^P P=0, \dots P-1$$

(5)

In Figure 7, a window of 3 × 3 size is used as an example, and the center pixel is taken as the threshold value and compared with the eight pixels in the neighborhood. If the center pixel is larger than the neighboring pixels, the pixel is assigned a value of 1; otherwise, it is 0, and an 8-bit binary number is generated, after which the binary numbers are arranged in the clockwise direction and the LBP value of the center pixel is calculated to be 19.

Feature extraction was performed on the collected aluminum blocks of different grades. Firstly, the BGR image was converted into an RGB image (src), which was in turn converted into a grayscale image (gray); the grayscale histogram was drawn, as were the histograms of the R, G, and B channels, and the LBP texture features were extracted from these.

Figure 8 shows the process of converting the input image (src) into RGB and grayscale, and then extracting LBP texture features. It can be seen that in the input images of the different grades, there were also obvious differences in the RGB images; there was a greater difference between the color shades of grade 1060 aluminum blocks and those of grades 5052 and 6061, and grade 5052 aluminum blocks were slightly darker than those of 6061. The histograms drawn from the grayscale images are shown in Figure 9.

From the above Figure 9, it can be seen that the grade 1060 aluminum block has a maximum value at a brightness of approximately 75, the grade 5052 and 6061 aluminum blocks have a maximum value at approximately 100–110; the extreme value of grade 1060 is larger than that of grades 5052 and 6061, and the extreme value of 12,000 can be used as the dividing line. At 190 brightness, the extreme value of 6061 is greater than 1000, much larger than those of 1060 and 5052, and this can be used as the dividing line between 5052 and 6061. It can be seen that the classification of the three grades can be realized through the grayscale images. The histograms of the R, G, and B channels of the RGB images are shown in Figure 10.

Specifically analyzing the differences between the R, B, and G color channels, the histograms show that in the R channel, the extreme values of 1060 and 6061 are both greater than 20,000, significantly higher than that of 5052, with 1060 slightly higher than 6061; in the B channel, the extreme values of 5052 and 6061 were both greater than 20,000, much greater than that of 1060, and the slopes of 5052 and 6061 were steeper while that of 1060 was relatively gentle; in the G channel, the extreme value of 1060 was larger than those of 5052 and 6061, and the extreme brightness of 1060 was near 75, while the extreme brightnesses of 5052 and 6061 were in the ranges of 110–114 and 124–126, respectively, with a clear demarcation line.

From the histograms of R, G, and B color channels, it can be seen that 1060 could be distinguished from 5052 and 6061 by the polar values of the G and B channels, while 5052 and 6061 could be distinguished from each other by the position interval of the polar luminance in the R and G channels. It can be seen that RGB images had obvious thresholds on each color channel, and therefore could achieve effective classification.

The RGB images were converted to HSV images, as shown in Figure 11, and the corresponding H, S, and V histograms were plotted, as shown in Figure 12 below.

The parameters of color in HSV are Hue (H), Saturation (S), and Value (V). As can be seen from the HSV chart, the images of 1060 differed from those of 5052 and 6061, and 5052 and 6061 looked slightly different.

According to the specific analysis of the HSV histogram, it can be seen that in the H channel, the extreme value of 1060 was much larger than those of 5052 and 6061, and the extreme value of 5052 was also slightly larger than that of 6061; in the S channel, the extreme value of 1060 was much smaller than the extreme values of 5052 and 6061, while the extreme value of 5052 was less than 20,000 and the extreme value of 6061 was larger than 20,000; in the V channel, the extreme values of 5052 and 6061 were much larger than that of 1060, the slopes were steeper, and the extreme value of 6061 was also slightly higher than that of 5052.

From the histograms of the H, S, and V color channels, it can be seen that 1060 could be distinguished from 5052 and 6061 by the extreme values in the H channel, while 5052 and 6061 could be distinguished by the extreme values in the S and V channels. However, the differentiation effect was not as good as that of the RGB images.

2.3. Data Set Creation and Pre-Processing

Upon collecting, classifying, and pre-processing the images of the three grades of aluminum sheets, the distribution of samples in the image datasets constructed in this paper is shown in

Table 1. Sample size in the database for different grades.

	1060	5052	6061	Total
RGB	259	343	357	959
HSV	497	497	518	1512
LBP	488	536	472	1496
Total	1244	1376	1347	3967

below.

Overall, 959 RGB images, 1512 HSV images, and 1496 LBP images were obtained, totaling 3967 images; these included 1244 aluminum blocks of grade 1060, 1376 aluminum blocks of grade 5052, and 1347 aluminum blocks of grade 6061. The training set and test set were classified based on the percentages of 80% and 20%.

In order to achieve fast training and convergence in a network model and improve the accuracy of the model, it is essential to pre-process the image database before performing network model training. In this paper, we used the Pytorch toolbox to process the images with random cropping and random flipping, convert the images to tensor format, and then normalize the tensor images. After pre-processing, the input image size of the network model was 224 × 224 × 3 (pixels × pixels × channels).

3. Construction of ResNet18 Classification Recognition Model

3.1. ResNet18 Network

As the depth of a neural network deepens, problems such as echelon disappearance and degradation occur, which leads to issues of difficult convergence and low accuracy of model training. A network model containing residual structure can largely avoid these problems; therefore, the ResNet18 model containing residual structure was used in this paper. The residual structure of this model is shown in Figure 13.

The solid line of residual structure a and the dashed line of residual structure b constitute the downsampling of the input. The residual block a indicates that the dimensional information is not changed, i.e., the length and width of the output feature matrix are the same as those of the input; the residual block b indicates that the dimensional information is changed, and the step size of the residual block in this paper is two, i.e., the length and width of the output feature matrix are half those of the input. The ResNet18 network with residual blocks increases the dimensionality of the output features and effectively avoids problems such as gradient disappearance and degradation.

The structure of the ResNet18 model used in this paper is shown in Figure 14.

The image in the ResNet18 model structure indicates an aluminum block image with an input size of 224 pixels × 224 pixels × 3 channels. In the neural network structure, conv is the convolutional layer, max pool is the maximum pooling layer, avg pool is the average pooling layer, and FC is the fully connected layer, such that the size of the convolutional kernel in 7 × 7conv, 64,/2 is 7 × 7, the number of channels is 64, and the step size is 2.

3.2. Dataset Selection

For the three different grades of aluminum blocks in this paper, a dataset of 3967 images of RGB, HSV, and LBP color channels was produced. In order to compare the impact of the three different color channels on the accuracy of the model, the resnet18 model network was first trained on the images of the above three color channels separately; the initial batch size was 16, the learning rate was set to 0.0001 and 200 iterations, and the accuracy of the resulting network model predictions and the training and testing accuracy for the RGB dataset are shown in Figure 15.

From the above figure, it can be seen that when LBP images were used for training, the accuracy of the test set could reach up to 97%, and when HSV images were used for training, the accuracy of the test set could reach up to 100%; however, the model’s accuracy with the LBP and HSV test sets fluctuated greatly, with the maximum fluctuation reaching more than 30%, and it could not converge. When RGB images were used for training, the network model converged quickly and was more stable with 100% accuracy. In Figure 15b, it can be observed that both the training and testing curves converged rapidly, and in most cases, the test accuracy was higher than that of the training set, which shows that there was no overfitting; meanwhile, the accuracy of both the training and test sets was high, which shows the superior performance of the model.

In summary, combining the model test accuracy curves, RGB images were chosen as the subsequent dataset for this paper.

4. Training and Hyperparameter Optimization of ResNet18 Model

4.1. Training Process

In this paper, we used the ResNet18 model to train on the RGB dataset. The dataset was loaded into the Pytorch deep learning framework; the input images were trained in batches according to the set batch size; the learning rate was adjusted by the Adam optimization algorithm; finally, the gradient of each iteration was calculated and updated. The number of iterations in this paper was 200. After 200 iterations, the accuracy and loss of the model were trained to convergence; the specific training process is shown in Figure 16.

4.2. Selection of Learning Rate

In this paper, the Adam optimizer was used for training. In the selection of a learning rate, when the learning rate is too low, it leads to slow convergence of the model; when the learning rate is too high, it leads to the inability of the model to converge, resulting in the loss function missing the optimal solution. In this paper, using the RGB dataset, the initial value of batch size was 16, and the learning rates of 0.0001, 0.00001, and 0.00005 were chosen for training to compare the effect of learning rate on the accuracy and loss of the model; the test results are shown in Figure 17.

As can be seen from the above figure, of the three different learning rates in the accuracy test, the highest could reach 100% accuracy, but with a learning rate of 0.0001 at 144 iterations there was a large fluctuation, and the accuracy rate dropped to 94.1%; with learning rates of 0.00001 and 0.00005, with the increase in the number of iterations, the fluctuation was lower and smaller and the accuracy became more and more stable. In general, the fluctuation in the accuracy rate was minor when the learning rate is 0.00001. After the loss of the three different learning rates rapidly decreased to 0.1, the loss value generally fluctuated above and below 0.05 as the number of iterations increased. Although all three also showed relatively large fluctuations, the loss values were kept below 0.1.

In summary, when the batch size was fixed at 16, 0.00001 was chosen as the learning rate of this paper by combining the accuracy of model testing and the convergence of training loss.

4.3. Selection of Batch Size

The batch size is also an essential hyperparameter for training the model. Although the impact of batch size on the model test accuracy and training loss is relatively small compared with that of the learning rate, the choice of the batch size is also critical to improving the stability of the model and reducing fluctuations. In this paper, we discuss the effects of three batch sizes on model test accuracy and training loss at a learning rate of 0.00001, i.e., the variation in model test accuracy and training loss for batch sizes of 8, 16, and 32, as shown in Figure 18.

As can be seen from the above figure, at a learning rate of 0.00001, all three different batch sizes converged quickly. When the batch size was 16, there were relatively large fluctuations in the early iterations while the results were relatively stable in the middle and late iterations, while for batch sizes 8 and 32, overall, there were relatively stable and shallow fluctuations. In the corresponding loss curves, it is obvious that the three different batch sizes had certain fluctuations after the rapid decline, but the overall trend was below 0.1; however, in the loss curve for batch size 32, the apparent fluctuations were more significant than in the other two, and the loss curve for batch size 8 had the smallest fluctuations and the most stable curve.

In summary, when the learning rate is fixed at 0.00001, combining the model test accuracy and the convergence of the training loss function, 8 was chosen as the batch size in this paper.

5. Results and Analysis

The aluminum block grades used in the experiments of this paper were 1060, 5052, and 6061, and the chemical compositions of the three grades are shown in

Table 2. Chemical composition ratios of the different grades of aluminum blocks (%).

	Cu≤	Si≤	Fe≤	Mn≤	Mg≤	Zn≤	Ti≤	Cr≤	Al
1060	0.05	0.25	0.35	0.3	0.03	0.1	0.03	0	99.6
5052	0.1	0.25	0.4	0.1	2.2~2.8	0.1	0	0.15~0.35	Residual
6061	0.15~0.4	0.4~0.8	0.7	0.15	0.8~1.2	0.25	0.15	0.04~0.35	Residual

.

The proportion of trace elements in the table is the maximum value of the element for each of the above three grades of aluminum used for data collection, classification, and production of the three datasets. Using the machine vision algorithm for the three datasets, images were compared, with the RGB data set as the subsequent dataset used. The ResNet18 network was used for training, with the initial learning rate set to 0.00001, the batch size to 8, and the number of iterations to 200; different network models were trained. A comparison of classification results from these different network models is shown in Figure 19 below.

As seen in the above figure, VGG16 could also reach 100% in classification and recognition accuracy, but fluctuated wildly and could not achieve convergence, and the overall trend of accuracy fluctuated from 90% to 100%. Compared with VGG16, SeNet had better performance in the accuracy and stability of classification and recognition and could achieve convergence, but the fluctuation range of SeNet was larger than that of ResNet18, which shows that the ResNet18 network was more stable.

The confusion matrix performed well in judging the merits of the classification models and classification effects, and it could clearly show the number of correct and incorrect identifications in aluminum block recognition. In this paper, we used the RGB dataset to train the ResNet18 model and visualize and analyze the recognition effect of the model; the results of the confusion matrix are shown in Figure 20.

Analysis of the confusion matrix shows that the recognized signs were distributed on the diagonal, which meant that all sign types could be successfully recognized.

6. Conclusions

In this paper, a method based on the combination of machine vision and deep learning is proposed to achieve the classification of different grades of nonferrous aluminum according to surface features and color space only, therefore no longer requiring complex instruments for element detection, effectively reducing detection cost and improving detection efficiency. With the help of the machine vision algorithm, we compared the surface features of RGB, HSV, and LBP images of three aluminum blocks with grades 1060, 5052, and 6061 and created the corresponding datasets, training with Pytorch as the deep learning framework, ResNet18 as the selected model, Relu as the chosen activation function, and with gradient update performed using the Adam optimization algorithm. The main conclusions are as follows.

(1)

The surface features of the aluminum blocks were extracted by machine vision, and results under different light sources and angles were compared. A blue light source and 45° angle lighting were selected, and the features of the acquired images in the RGB and HSV color spaces were compared. It could be seen that 1060 significantly differed from 5052 and 6061, and 5052 and 6061 could also be distinguished by setting the threshold value.

(2)

Datasets of RGB, HSV, and LBP images of the different grades of aluminum were created; these three datasets were selected to train the model using ResNet18, comparing the accuracy of the three datasets for aluminum classification and recognition, and ultimately RGB images were selected as the best dataset.

(3)

Using the RGB dataset to train the ResNet18 model with a learning rate of 0.00001 and a batch size of 8, the classification and recognition rate for different grades of aluminum could reach 100%; comparing the training effects of different network models, ResNet18 had significant advantages in model stability and accuracy, compared with the traditional use of spectra for classification and recognition of different grades of similar materials. This paper achieved faster, higher-volume, and lower-cost classification and recognition of different grades of aluminum.