1. Introduction

2. Related Works

2.1. Computer Vision-Based Defect Detection

2.1.1. Limitations of Deep Learning Techniques Based on Image Classification

Recently, deep learning technology has made great progress and is being used in various fields. In particular, deep learning models such as CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network) are also attracting attention in the field of image classification. Additionally, the recent emergence of new architectures such as vision transformer (ViT) expands image processing possibilities, providing CNN-like performance. However, while this approach shows excellent performance, it still has several limitations.

Figure 2. This figure shows how to amplify and reduce data to address imbalances in a dataset, with over sampling on the left and under sampling on the right.

Figure 2. This figure shows how to amplify and reduce data to address imbalances in a dataset, with over sampling on the left and under sampling on the right.

First, there is a class imbalance problem. In most datasets, the imbalance between majority and minority classes has a negative impact on the performance of deep learning models [15]. In particular, when class imbalance is severe, feature learning of minority classes is not performed properly, and the model tends to perform excellently only in predictions for majority classes [16]. In a recent study, the performance degradation of deep learning models for various class imbalance problems was analyzed, and it was found that the larger the imbalance, the greater the tendency for model accuracy to deteriorate [17,18]. To solve this problem, use random sampling technology to under sample the majority category data to fit the minority category data, or use a technology called GAN (Generative Adversarial Nets) to oversample the minority category data. Even if these techniques are used, accurate measurement is difficult [19]. Second, deep learning models such as CNN have high computational complexity and require a lot of resources during the learning and inference process [18]. Especially in a real-time environment, this requires research to optimize the computational load for real-time product quality evaluation using multiple frame rate cameras and various deep learning models [20,21]. CNN can extract various spatial features of an image through multiple layers of convolution, but its computational cost is very high due to its structural characteristics. This can cause difficulties in applying deep learning models when real-time processing is required in an actual industrial environment, and in order to solve these problems, several studies have argued that lightweight network design or hardware acceleration techniques are needed. . Third, using deep learning models in a manufacturing environment is subject to various limitations. Due to the nature of the manufacturing industry, the model must be retrained every time a new product is manufactured in addition to an existing product, which inevitably requires significant time and cost during the data collection and labeling process. Additionally, if the manufacturing facility or environment changes, the measurement environment must be rebuilt and the model must be retrained, which can have a negative impact on operational efficiency.

To solve these limitations, this paper uses computer vision techniques such as SSIM, MSE, PSNR, and SIFT. and compare the image data. The goal is to compare good and defective products without the need for a predefined learning data set.

2.1.2. Application of Vision Algorithm in Product Classification

The development of vision algorithms is currently making an important contribution to the detection and classification of product defects and quality control in the manufacturing industry, and continues to develop in the direction of increasing real-time processing capabilities and accuracy. Through this, it plays an important role in maximizing productivity and quality control efficiency in the manufacturing process. When applying vision algorithms in product classification, commonly used image comparison techniques such as SSIM, MSE, and PSNR are used. This is a technology that evaluates structural similarity and pixel differences between the same images. Although these techniques are useful for basic image quality evaluation, they have the limitation of not being sufficiently robust against rotation or size changes.

Figure 3. This figure shows a comparative analysis of the performance of the SIFT, SURF, and hybrid algorithms. On the left is the performance metric for scale change, and on the right is the performance metric for rotation.

Figure 3. This figure shows a comparative analysis of the performance of the SIFT, SURF, and hybrid algorithms. On the left is the performance metric for scale change, and on the right is the performance metric for rotation.

In contrast, feature-based algorithms such as SIFT can extract feature points robustly even when image rotation or size changes, enabling accurate defect detection [22,23]. SIFT has the advantage of being able to extract feature points stably despite rotation or size changes within the image [24,25]. According to research results, when testing SURF (Speed-Up Robust Features), SIFT, and Hybrid technologies, hybrid was superior in scale changes, but SIFT was found to provide more accurate detection when detecting features such as rotation invariance [26]. Based on this, this paper aims to develop a model that increases the recognition rate of products and enables more accurate defect detection by utilizing the SIFT technique to overcome the limitations of existing techniques.

2.2. Research on Body Movement Detection Based on Body Skeletal Structure

2.2.1. Latest Research Trends in Body Movement Detection

Body movement detection is advancing as a technology that leverages deep learning, a field within machine learning, to identify human movement, detect hazards, and recognize specific actions. In HAR (Human Activity Recognition), deep learning models play a crucial role in accurately recognizing various movements by analyzing video and sensor data. The development of HAR technology primarily involves the integration of diverse deep learning architectures [27]. Applications such as real-time hand gesture recognition are made possible by these advancements, and research has even extended to gesture recognition based on body skeletal structure [28]. These studies highlight the impressive capabilities of deep learning in tracking and recognizing body movements in real time and suggest which approaches may be more effective in particular situations by comparing different architectures.

Figure 4. This figure is an image of the YOLO V8 architecture to illustrate the overall system. [29].

Figure 4. This figure is an image of the YOLO V8 architecture to illustrate the overall system. [29].

One such architecture, the YOLO v8 pose model, incorporates advanced features for pose estimation, offering enhanced speed and accuracy over previous versions. It utilizes the new C2f module, which allows for lightweight computation and efficient feature reuse, optimizing the model for dense and complex tasks like pose estimation. Unlike previous YOLO models, YOLO v8 is anchor-free, simplifying the detection process. This architecture includes a streamlined backbone—the core neural network responsible for feature extraction—enabling more efficient keypoint detection. These improvements make YOLO v8 particularly well-suited for real-time applications such as human motion tracking [30].

Figure 5. This figure shows a graph comparing the performance of the YOLO object detection model. [31].

Figure 5. This figure shows a graph comparing the performance of the YOLO object detection model. [31].

The following summarizes the performance evolution across YOLO versions from v1 to v8. YOLOv1 introduced real-time object detection with a single network pass but struggled with small object detection. YOLOv2 improved accuracy with anchor boxes and batch normalization, while YOLOv3 enhanced multi-scale object detection by adding the Darknet-53 backbone and multi-scale prediction capabilities. YOLOv4 achieved a balance between speed and accuracy through CSPDarknet-53, SPP, and PAN modules, and YOLOv5 increased user-friendliness and learning efficiency by transitioning to PyTorch. Building on this, YOLOv6 introduced RepVGG blocks and a decoupled head structure for better performance, and YOLOv7 achieved the highest accuracy and speed by maximizing computational efficiency with E-ELAN and RepConv [32]. The latest version, YOLO v8, represents the most advanced iteration, enhancing speed and accuracy with an anchor-free design and the C2f module. For these reasons, this study selects YOLO v8 as the basis for further research and experimentation.

3. Methods and Result

3.1. Image-Based Product Quality Inspection and Feature Matching Techniques

3.1.1. Rotation and Scale Invariance in Image Comparison Using SIFT Algorithm

As shown in Figure 6, after loading two images, the SIFT (Scale-Invariant Feature Transform) algorithm is used to detect keypoints within the images and calculate descriptors for those keypoints. SIFT is an algorithm designed to detect keypoints that are invariant to scale, rotation, and illumination changes, enabling it to reliably find consistent features even under various transformations [38]. The first step of the SIFT algorithm is to locate keypoints in the scale space. This is achieved by using a Gaussian filter to process the image at different scales, allowing the detection of important keypoints by identifying extrema (maxima and minima) at each scale. The process of finding keypoints in the scale space, where a Gaussian blur is applied, utilizes the Difference of Gaussian (DoG) method [41]:

$$D(x,y,\sigma )=\left(G\right(x,y,k\sigma )-G(x,y,\sigma \left)\right)\ast I(x,y)$$

(1)

Here, G(x,y,σ) represents the image with Gaussian blur applied at scale σ, I(x,y) is the original image, and k denotes the scale factor. In this process, the differences between images at each scale are computed to identify extrema (maxima/minima), which are detected as keypoints [39,40,41]. By calculating the gradient magnitude and orientation of the pixels surrounding each keypoint, a principal orientation is assigned to each keypoint to ensure rotational invariance. The gradient magnitude m(x,y) and orientation θ(x,y) are computed using the following equations:

$$\mathit{m}\left(\mathit{x},\mathit{y}\right)=\sqrt{{\mathit{L}(\mathit{x}+1,\mathit{y})}^2-\mathit{L}(\mathit{x}-1,\mathit{y}))+(\mathit{L}(\mathit{x},\mathit{y}+1)-{\mathit{L}(\mathit{x},\mathit{y}-1))}^2}$$

(2)

$$\theta \left(x,y\right)={\mathit{tan}}^{-1}({\displaystyle \frac{L(x,y+1)-L(x,y-1)}{L\left(x+1,y\right)-L(x-1,y)}}$$

(3)

Here, L(x,y) represents the intensity of the image, m(x,y) is the gradient magnitude at that position, and θ(x,y) is the gradient orientation. Based on the calculated orientation information, a principal direction is assigned to each keypoint, enabling the keypoints to maintain rotational invariance [39,40,41]. To proceed with the matching process between the two images, the Euclidean distance between each descriptor is calculated to assess their similarity. The distance between two descriptors, p and q, is defined as follows:

$$d\left(p,q\right)=\sqrt{\sum _{i=1}^{128}{\left(p_i-q_i\right)}^2}$$

(4)

Here, $p_i$ and $q_i$ represent the iii-th components of the two descriptors, respectively. The smaller the Euclidean distance, the more similar the two descriptors are considered, enabling the matching of keypoints between the two images. Once the matching is completed, the rotation angle between the two images can be estimated, and this angle is calculated using the following equation:

$$\theta ={\mathit{tan}}^{-1}\left({\displaystyle \frac{y_2-y_1}{x_2-x_1}}\right)$$

(5)

Here, ($x_1$,$y_1$) and ($x_2$,$y_2$) represent the coordinates of the matched keypoint pairs in the two images, respectively. Using this equation, the rotation angle between the two images can be estimated, allowing for rotation correction and image restoration based on this information. Additionally, by calculating the distance between the matched keypoints, the scale variation between the two images can be assessed, enabling the identification of whether the image has been scaled up or down.

Figure 6. This figure shows an example of extracting descriptors from two input images using the SIFT algorithm, where one image serves as the reference. The algorithm restores the original rotation of the second image based on the reference, illustrating the process of image alignment.

Figure 6. This figure shows an example of extracting descriptors from two input images using the SIFT algorithm, where one image serves as the reference. The algorithm restores the original rotation of the second image based on the reference, illustrating the process of image alignment.

3.1.2. Defect Detection Using a Combined SSIM, PSNR, and MSE Evaluation

Figure 7 shows the difference images between A product and B product obtained through pre-processing with SIFT, as well as the difference image between the two. To quantitatively evaluate the similarity between the two images, SSIM (Structural Similarity Index), PSNR (Peak Signal-to-Noise Ratio), and MSE (Mean Squared Error) were applied. These metrics are used to accurately compare and detect defects between normal and defective products, enabling efficient identification of product defects. The restored images were evaluated using SSIM, PSNR, and MSE to quantitatively assess the similarity between the two images, with the formulas for each method shown below [42,43,44]:

Table 1. Formulas and Descriptions for SSIM, MSE, and PSNR Metrics.

Table 1. Formulas and Descriptions for SSIM, MSE, and PSNR Metrics.

Metric	Formula
$\mathrm{S}\mathrm{S}\mathrm{I}\mathrm{M}$	${\displaystyle \frac{(2{\mu }_x{\mu }_y+C_1)(2{\sigma }_{xy}+C_2)}{({\mu }_x^2+{\mu }_y^2+C_1)({\sigma }_x^2+{\sigma }_y^2+C_2)}}$ (6) ${\mu }_x$ and ${\mu }_y$: The mean brightness values of the two images ${\sigma }_x^2$ and ${\sigma }_y^2$: The variance of each image ${\sigma }_{xy}$: The covariance between the two images $C_1$ and $C_2$: Constants for stability
$\mathrm{M}\mathrm{S}\mathrm{E}$	${\displaystyle \frac{1}{mn}}{\sum }_{i=1}^m{\sum }_{j=1}^n{\left[I_1\left(i,j\right)-I_2(i,j)\right]}^2$ (7) $I_1(i,j)$ and $I_2(i,j)$: The pixel values at the $i,j$ coordinates of the two images $m$ and $n$: The dimensions of the image
$\mathrm{P}\mathrm{S}\mathrm{N}\mathrm{R}$	$10{\mathit{log}}_{10}\left({\displaystyle \frac{MA{X}_I^2}{MSE}}\right)$(8) $MA{X}_I$: The maximum possible pixel value in the image $MSE$: The Mean Squared Error between the two images

By applying SSIM, PSNR, and MSE to the restored images, a quantitative evaluation was performed based on structural similarity, signal-to-noise ratio, and mean squared error between the two images. SSIM measures the structural similarity of the images, PSNR assesses the signal-to-noise ratio, and MSE analyzes the detailed pixel differences, enabling an overall evaluation of similarity. These metrics can vary according to specific criteria set by the user and may be interpreted differently depending on the goals, application area, and quality requirements of the image processing task. Therefore, users can establish standards for each metric according to the project’s objectives and requirements and evaluate image similarity based on these standards. For example, in manufacturing processes where defect detection is critical, even minor defects may significantly impact results. Thus, an SSIM value of 0.95 or higher could indicate an acceptable product, and a PSNR value of 40 dB or higher could suggest good quality [42]. In the case of MSE, a lower value signifies fewer differences between the two images. Therefore, to ensure consistent interpretation within $S_{combined}$, the inverse of MSE is used in calculations. This approach allows for a more intuitive interpretation, where a higher $S_{combined}$ score indicates greater similarity between the images. Finally, the similarity score $S_{combined}$, reflecting the weights of SSIM, PSNR, and MSE, is calculated as follows:

$$S_{combined}={\omega }_{SSIM}\ast S_{SSIM}+{\omega }_{PSNR}\ast S_{PSNR}+{\omega }_{MSE}\ast {\displaystyle \frac{1}{S_{MSE}}}$$

(9)

The weights ${\omega }_{SSIM}$ ,${\omega }_{PSNR}$,${\omega }_{MSE}$ in the $S_{combined}$ formula sum to 1 and are adjusted based on the characteristics that each metric evaluates. This adjustment is not just about evaluating the images, but also about focusing on the specific defects and characteristics of the product. First, if the overall appearance or structure of the product is important, the weight of the SSIM metric, ${\omega }_{SSIM}$ , is increased. SSIM evaluates the structural similarity of the image, focusing on structural elements such as patterns, edges, and textures of the product. This is suitable in situations where structural defects are critical, such as in the bending of metal products or the consistency of patterns in textiles. Second, when noise or distortion on the product’s surface is the focus, the weight of the PSNR metric, ${\omega }_{PSNR}$, is increased. PSNR plays a significant role in examining surface scratches or the finishing of lens surfaces. Third, when fine pixel-level defects are of particular importance, the weight of the MSE metric, ${\omega }_{MSE}$ , is increased. MSE precisely calculates the differences between pixels, making it ideal for processes that need to detect very small defects. In conclusion, the weights in the $S_{combined}$ formula are set according to the product’s characteristics and the type of defects being emphasized. By adjusting the weights according to the specific features of each metric, the efficiency and accuracy of defect detection can be improved.

3.2. Product Counting Algorithm Using Camera-Based Skeleton Tracking and Body Part Detection

3.2.1. Classification of Counting Classes Based on Body Part Detection

In this study, we developed a product counting system that applies the YOLO v8 Pose algorithm to detect the human body and correct errors that occur when a person steps onto the load cell [45,46]. The core of the research lies in using a camera and algorithm to detect human body parts in real time, distinguishing factors that affect weight data measured by the load cell, and reducing counting errors. Body parts that influence load cell weight measurements were categorized into four classes, and the load cell’s weight measurement actions were controlled according to each class.

• Full Upper Body Detection: When the full upper body is detected within the load cell area, weight measurement is temporarily paused. This is because the structure, in which the camera views the load cell from above, may result in the lower body or other body parts not being detected. When the full upper body is detected, weight measurement is paused to eliminate the influence of the body on the load cell, and the measurement resumes once the upper body moves away from the load cell.
• Partial Upper Body Detection: In partial upper body detection, only a part of the upper body is detected within the load cell area. During this time, the load cell continuously reads weight data in real time, and only when the weight change meets the stabilization value is it considered valid. Weight changes less than 0.5 kg are regarded as insignificant fluctuations and are not included in the count. Therefore, when partial upper body detection occurs, the load cell measurement continues, but small weight changes are ignored, and only meaningful changes are reflected in the count.
• Lower Body Detection: When the lower body is detected within the load cell area, weight measurement is paused. The weight of the lower body directly affects the load cell, so weight changes are not measured while the lower body is detected. Measurement resumes when the lower body moves away from the load cell area.
• No Detection: If the camera does not detect any part of the body over the load cell, the load cell continuously measures weight changes in real time and counts the product based on the weight variations. In the no detection state, the load cell counting process proceeds normally, and the measured weight changes are used to calculate the product count.

3.2.2. Overall Flowchart of the Product Counting Algorithm

Figure 9. This figure shows the step-by-step process flow of a product counting system.

Figure 9. This figure shows the step-by-step process flow of a product counting system.

3.2.3. Overall Flowchart of the Product Counting Algorithm

Figure 10. This figure shows the process of unit weight calculation using the image differencing technique.

Figure 10. This figure shows the process of unit weight calculation using the image differencing technique.

The load cell used in this system can measure weights up to 2000kg and provides a resolution of 500g. While higher resolution increases the precision of the load cell system, it also raises the cost. Load cells with high resolution enable precise measurements, but in many cases, the level of precision exceeds what is required in industrial settings. In particular, the products measured in this system mostly weigh over 1kg, so a resolution of 500g is sufficient for accurate weight counting. This study focuses on developing a method to count products using weight increments of 500g, based on these conditions.

In this system, the image differencing technique is used to determine the unit weight of products [47,48,49]. The image differencing method detects the number of products placed on the load cell and calculates the unit weight based on the total weight of the products. During the initial setup, a certain number of products are placed on the load cell, and the unit weight is calculated by using the number of products detected through image differencing and the total weight measured by the load cell. This technique enables accurate detection of the number of products, minimizing weight measurement errors caused by product placement. To enhance the reliability of the image differencing method, adjustments were made to compensate for external factors such as camera angle, lighting, and background noise. This ensures minimal impact from environmental changes on product detection, improving the accuracy of the product counting process.

$$W_{unit}={\displaystyle \frac{W_{Total}}{N_{product}}}$$

(10)

Here, $W_{unit}$ represents the unit weight of the product, $W_{Total}$ is the total weight measured by the load cell, and $N_{product}$ is the number of products detected using the image differencing technique. Once the initial unit weight is established, the number of products is calculated based on the change in weight measured by the load cell, following these steps:

• Weight Change Calculation: When products are added or removed, the weight change is calculated by determining the difference between the current and previous weights measured by the load cell. The weight change must exceed a certain percentage of the unit weight (e.g., 0.5) to be considered valid. This prevents counting errors caused by minor weight fluctuations.

$$∆W=W_{current}-W_{previous}$$

(11)

• Stabilization Process: To improve counting accuracy, the system detects the point at which the weight change stabilizes. Based on the number of data points received per second, $N_{DATA }$, if the same weight change is detected over a certain number of consecutive readings, the weight is considered stabilized, and the counting is performed. This helps to reduce errors caused by temporary weight fluctuations or noise.
• Precise Decimal Handling: The weight change is calculated with precision down to the decimal places, and rounding or truncation is applied only at the time of counting. This minimizes counting errors that may occur when multiple products are loaded simultaneously. For example, if the weight change appears in increments of 0.5, it is rounded up and reflected in the final count.
• Counting Execution: Once the stabilization process is complete, the number of products is calculated by dividing the weight change by the unit weight. During this process, decimal values are carefully handled, and if the first decimal place is 0.5, rounding up or down is applied to ensure the accuracy of the count.
• Stabilization Process: To enhance counting accuracy, the system detects when the weight change stabilizes. Based on the number of data points received per second, $N_{DATA}$, if the same weight change is detected consistently over a certain number of readings, the weight is considered stabilized, and counting is performed. This helps reduce errors caused by temporary weight fluctuations or noise. Here, ${\mathrm{C}}_{\mathrm{p}\mathrm{r}\mathrm{o}\mathrm{d}\mathrm{u}\mathrm{c}\mathrm{t}}$ represents the number of products, ∆W is the weight change, and $W_{unit}$ is the unit weight.

  $$C_{product}={\displaystyle \frac{∆W}{W_{unit}}}$$

  (12)

• Count Error Correction: In this system, only products weighing 1kg or more are counted. If the weight change is less than 1kg, it is excluded from the count. Specifically, weight changes of 0.5kg or less are considered minor fluctuations and are disregarded in the count results. This prevents errors caused by small weight changes and ensures that only the actual weight changes of the products are accurately reflected in the count.

Figure 11. This is an image of the test result showing the product counting accuracy.

Figure 11. This is an image of the test result showing the product counting accuracy.

Table 2. Product Counting Accuracy Test Results.

Table 2. Product Counting Accuracy Test Results.

Test	Product	Product Weight	Actual Quantity	Estimated Quantity	Accuracy(%)
1	A	6.00	184	183	99.46
2	B	43.58	24	25	95.83
3	C	3.1	254	254	100
4	D	3.93	92	91	98.91
5	E	11.80	42	42	100
6	F	4.27	66	67	98.48
7	G	14.44	35	35	100
8	H	4.23	111	111	100
9	I	1.35	50	50	100
10	J	2.3	30	30	100

$$Overall Accuracy\left(\%\right)={\displaystyle \frac{\sum Accuracy for Each Test}{Number of Tests}}$$

(13)

The experiment was conducted in collaboration with a manufacturer of automotive parts, and the test involved counting 10 products ranging from 1kg to 45kg. An accuracy of 99.268% was achieved in the product counting test.

4. Discussion

5. Conclusions

References

1. Reyes Domínguez, D.; Infante Abreu, M. B.; Parv, A. L. Main Trend Topics on Industry 4.0 in the Manufacturing Sector: A Bibliometric Review. Appl. Sci. 2024, 14, 6450. [Google Scholar] [CrossRef]
2. Wang, Z.; Zhao, L.; Li, H.; Xue, X.; Liu, H. Research on a Metal Surface Defect Detection Algorithm Based on DSL-YOLO. Sensors (Basel) 2024, 24, 6268. [Google Scholar] [CrossRef] [PubMed]
3. Ahmmed, M. S.; Isanaka, S. P.; Liou, F. Promoting Synergies to Improve Manufacturing Efficiency in Industrial Material Processing: A Systematic Review of Industry 4.0 and AI. Machines 2024, 12, 681. [Google Scholar] [CrossRef]
4. Lin, B. H.; Chen, J. C.; Lien, J. J. J. Defect Inspection Using Modified YoloV4 on a Stitched Image of a Spinning Tool. Sensors 2023, 23, 4476. [Google Scholar] [CrossRef]
5. Sara, U.; Akter, M.; Uddin, M. S. Image Quality Assessment through FSIM, SSIM, MSE and PSNR—A Comparative Study. J. Comput. Commun. 2019, 7, 8–18. [Google Scholar] [CrossRef]
6. Zhou, L.; Zhang, L.; Konz, N. Computer Vision Techniques in Manufacturing. IEEE Trans. Syst. Man Cybern. Syst. 2022, 53, 105–117. [Google Scholar] [CrossRef]
7. Sabilla, I. A.; Meirisdiana, M.; Sunaryono, D.; Husni, M. Best Ratio Size of Image in Steganography Using Portable Document Format with Evaluation RMSE, PSNR, and SSIM. In Proceedings of the 2021 4th International Conference of Computer and Informatics Engineering (IC2IE); IEEE: September, 2021; pp. 289–294. [Google Scholar]
8. Prasannakumar, A.; Mishra, D. Deep Efficient Data Association for Multi-Object Tracking: Augmented with SSIM-Based Ambiguity Elimination. J. Imaging 2024, 10. [Google Scholar] [CrossRef]
9. Shahsavarani, S.; Lopez, F.; Ibarra-Castanedo, C.; Maldague, X. P. Advanced Image Stitching Method for Dual-Sensor Inspection. Sensors 2024, 24, 3778. [Google Scholar] [CrossRef]
10. Zhang, H.; Zheng, R.; Zhang, W.; Shao, J.; Miao, J. An Improved SIFT Underwater Image Stitching Method. Appl. Sci. 2023, 13, 12251. [Google Scholar] [CrossRef]
11. Tsourounis, D.; Kastaniotis, D.; Theoharatos, C.; Kazantzidis, A.; Economou, G. SIFT-CNN: When Convolutional Neural Networks Meet Dense SIFT Descriptors for Image and Sequence Classification. J. Imaging 2022, 8, 256. [Google Scholar] [CrossRef]
12. Bansal, M.; Kumar, M.; Kumar, M. 2D Object Recognition: A Comparative Analysis of SIFT, SURF and ORB Feature Descriptors. Multimed. Tools Appl. 2021, 80, 18839–18857. [Google Scholar] [CrossRef]
13. Lozano-Vázquez, L. V.; Miura, J.; Rosales-Silva, A. J.; Luviano-Juárez, A.; Mújica-Vargas, D. Analysis of Different Image Enhancement and Feature Extraction Methods. Mathematics 2022, 10, 2407. [Google Scholar] [CrossRef]
14. Bellavia, F.; Colombo, C. Is There Anything New to Say About SIFT Matching? Int. J. Comput. Vis. 2020, 128, 1847–1866. [Google Scholar] [CrossRef]
15. Talaei Khoei, T.; Ould Slimane, H.; Kaabouch, N. Deep Learning: Systematic Review, Models, Challenges, and Research Directions. Neural Comput. Appl. 2023, 35, 23103–23124. [Google Scholar] [CrossRef]
16. Ghosh, K.; Bellinger, C.; Corizzo, R.; Branco, P.; Krawczyk, B.; Japkowicz, N. The Class Imbalance Problem in Deep Learning. Mach. Learn. 2024, 113, 4845–4901. [Google Scholar] [CrossRef]
17. Hütten, N.; Alves Gomes, M.; Hölken, F.; Andricevic, K.; Meyes, R.; Meisen, T. Deep Learning for Automated Visual Inspection in Manufacturing and Maintenance: A Survey of Open-Access Papers. Appl. Syst. Innov. 2024, 7, 11. [Google Scholar] [CrossRef]
18. Zhong, X.; Zhu, J.; Liu, W.; Hu, C.; Deng, Y.; Wu, Z. An Overview of Image Generation of Industrial Surface Defects. Sensors 2023, 23, 8160. [Google Scholar] [CrossRef]
19. Kumar, V.; Lalotra, G. S.; Sasikala, P.; Rajput, D. S.; Kaluri, R.; Lakshmanna, K.; Uddin, M. Addressing Binary Classification Over Class Imbalanced Clinical Datasets Using Computationally Intelligent Techniques. Healthcare; 2022; 10, p. 1293. [Google Scholar]
20. Wibowo, A.; Setiawan, J. D.; Afrisal, H.; Mertha, A. A. S. M. M. J.; Santosa, S. P.; Wisnu, K. B.; Caesarendra, W. Optimization of Computational Resources for Real-Time Product Quality Assessment Using Deep Learning and Multiple High Frame Rate Camera Sensors. Appl. Syst. Innov. 2023, 6, 25. [Google Scholar] [CrossRef]
21. Archana, R.; Jeevaraj, P. E. Deep Learning Models for Digital Image Processing: A Review. Artif. Intell. Rev. 2024, 57, 11. [Google Scholar] [CrossRef]
22. Burger, W.; Burge, M. J. Scale-Invariant Feature Transform (SIFT). In Digital Image Processing: An Algorithmic Introduction; Springer International Publishing: Cham, 2022. [Google Scholar]
23. Lowe, D. G. Distinctive Image Features from Scale-Invariant Keypoints. Int. J. Comput. Vis. 2004, 60, 91–110. [Google Scholar] [CrossRef]
24. Acharya, K. A.; Venkatesh Babu, R.; Vadhiyar, S. S. A Real-Time Implementation of SIFT Using GPU. J. Real-Time Image Process. 2018, 14, 267–277. [Google Scholar] [CrossRef]
25. Karami, E.; Prasad, S.; Shehata, M. Image Matching Using SIFT, SURF, BRIEF and ORB: Performance Comparison for Distorted Images. arXiv 2017, arXiv:1710.02726. [Google Scholar]
26. Güzel, M. S. A Hybrid Feature Extractor Using Fast Hessian Detector and SIFT. Technologies 2015, 3, 103–110. [Google Scholar] [CrossRef]
27. Kumar, P.; Chauhan, S.; Awasthi, L. K. Human Activity Recognition (HAR) Using Deep Learning: Review, Methodologies, Progress and Future Research Directions. Arch. Comput. Methods Eng. 2024, 31, 179–219. [Google Scholar] [CrossRef]
28. Aggarwal, A.; Bhutani, N.; Kapur, R.; Dhand, G.; Sheoran, K. Real-Time Hand Gesture Recognition Using Multiple Deep Learning Architectures. Signal Image Video Process. 2023, 17, 3963–3971. [Google Scholar] [CrossRef]
29. Ultralytics. https://github.com/ultralytics/ultralytics/issues/189 (accessed 2023-11-11).
30. Talaat, F. M.; ZainEldin, H. An Improved Fire Detection Approach Based on YOLO-v8 for Smart Cities. Neural Comput. Appl. 2023, 35, 20939–20954. [Google Scholar] [CrossRef]
31. Ultralytics. YOLOv8—Ultralytics YOLOv8 Documentation. https://docs.ultralytics.com/models/yolov8/ (accessed 2023-11-11).
32. Terven, J.; Córdova-Esparza, D. M.; Romero-González, J. A. A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS. Mach. Learn. Knowl. Extr. 2023, 5, 1680–1716. [Google Scholar] [CrossRef]
33. Barton, M.; Budjac, R.; Tanuska, P.; Sladek, I.; Nemeth, M. Advancing Small and Medium-Sized Enterprise Manufacturing: Framework for IoT-Based Data Collection in Industry 4. 0 Concept. Electronics 2024, 13(13), 2485. [Google Scholar] [CrossRef]
34. Zhen, L.; Li, H. A Literature Review of Smart Warehouse Operations Management. Front. Eng. Manag. 2022, 9, 31–55. [Google Scholar] [CrossRef]
35. Ryalat, M.; Franco, E.; Elmoaqet, H.; Almtireen, N.; Alrefai, G. The Integration of Advanced Mechatronic Systems into Industry 4. 0 for Smart Manufacturing. Sustainability 2024, 16, 8504. [Google Scholar]
36. Albayrak Ünal, Ö.; Erkayman, B.; Usanmaz, B. Applications of Artificial Intelligence in Inventory Management: A Systematic Review of the Literature. Arch. Comput. Methods Eng. 2023, 30, 2605–2625. [Google Scholar] [CrossRef]
37. Rocha, J. G.; Couto, C.; Correia, J. H. Smart Load Cells: An Industrial Application. Sens. Actuators, A 2000, 85, (1–3). [Google Scholar] [CrossRef]
38. Hossein-Nejad, Z.; Agahi, H.; Mahmoodzadeh, A. Image Matching Based on the Adaptive Redundant Keypoint Elimination Method in the SIFT Algorithm. Pattern Anal. Appl. 2021, 24, 669–683. [Google Scholar] [CrossRef]
39. Hu, X.; Tang, Y.; Zhang, Z. Video Object Matching Based on SIFT Algorithm. In 2008 International Conference on Neural Networks and Signal Processing; IEEE, 2008; pp 412–415. [CrossRef]
40. Alhwarin, F.; Wang, C.; Ristić-Durrant, D.; Gräser, A. Improved SIFT-Features Matching for Object Recognition. In Visions of Computer Science-BCS International Academic Conference; BCS Learning & Development, 2008; pp 165–176.
41. Zhou, H.; Yuan, Y.; Shi, C. Object Tracking Using SIFT Features and Mean Shift. Comput. Vis. Image Underst. 2009, 113, 345–352. [Google Scholar] [CrossRef]
42. Setiadi, D. R. I. M. PSNR vs SSIM: Imperceptibility Quality Assessment for Image Steganography. Multimedia Tools Appl. 2021, 80, 8423–8444. [Google Scholar] [CrossRef]
43. Hore, A.; Ziou, D. Image Quality Metrics: PSNR vs. SSIM. In 2010 20th International Conference on Pattern Recognition; IEEE, 2010; pp 2366–2369. [CrossRef]
44. Sara, U.; Akter, M.; Uddin, M. S. Image Quality Assessment through FSIM, SSIM, MSE, and PSNR—A Comparative Study. J. Comput. Commun. 2019, 7, 8–18. [Google Scholar] [CrossRef]
45. Terven, J.; Córdova-Esparza, D. M.; Romero-González, J. A. A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS. Mach. Learn. Knowl. Extr. 2023, 5, 1680–1716. [Google Scholar] [CrossRef]
46. Jocher, G.; Chaurasia, A.; Qiu, J. YOLO by Ultralytics; GitHub, 2023. Available online: https://github.com/ultralytics/ultralytics (accessed on January 12, 2023).
47. Piccardi, M. Background Subtraction Techniques: A Review. In 2004 IEEE International Conference on Systems, Man and Cybernetics; IEEE, 2004; Vol. 4, pp 3099–3104. [CrossRef]
48. Kalsotra, R.; Arora, S. Background Subtraction for Moving Object Detection: Explorations of Recent Developments and Challenges. Vis. Comput. 2022, 38, 4151–4178. [Google Scholar] [CrossRef]
49. Benezeth, Y.; Jodoin, P. M.; Emile, B.; Laurent, H.; Rosenberger, C. Comparative Study of Background Subtraction Algorithms. J. Electron. Imaging 2010, 19, 033003. [Google Scholar] [CrossRef]