The method of filling the dataset of mobile military objects was studied. In the future, objects captured by a video camera mounted on a moving object will be investigated. The software used is YOLO v8, which allows you to track moving objects that fall into the video from the video camera. We use artificial intelligence methods to recognize military equipment. Improvements in object recognition can be achieved by contour analysis, pattern comparison, and point-by-point comparison. In future works, we will develop recognition of these theories. The metrics used to evaluate object recognition are shown. A method of filling the dataset and creating a classifier is proposed. Shown are graphs, the results of moving object recognition in Yolo8x.
intelligence based on deep learning
technology, convolutional neural networks [
In the field of computer vision, machine
vision is involved in the automatic recognition
of moving objects, and the detection of
obstacles during the movement of objects [
The YOLO series has been updated to YOLO v8
[8]. To further improve the performance of
existing object detection algorithms, many
scientists have begun to study them. In
particular, [
In an era marked by rapid advancements in Artificial Intelligence (AI) and computer vision, the convergence of these technologies has ignited revolutionary progressions with profound implications. The domain of military applications stands at the forefront of harnessing AI’s potential, particularly within the sphere of computer vision. This article embarks on a critical exploration of this aspect, delving into the importance of training AI models, specifically focusing on the You Only Look Once (YOLO) model, for the recognition of military objects, with a specific emphasis on tanks.
As the velocity of technological innovation intensifies, AI and computer vision have evolved into crucial instruments that reshape our perceptions and interactions with the environment. In military contexts, the incorporation of computer vision capabilities holds paramount significance, providing heightened situational awareness, facilitating strategic decision-making, and optimizing overall operational efficiency. The rapid and accurate identification of military objects, such as tanks, is imperative in environments characterized by dynamism and unpredictability.
Amidst these technological advancements, recent geopolitical events, and the war in Ukraine, accentuate the urgent requirement for advanced AI models capable of detecting military assets like tanks. The deployment of AI in conflict zones confers a strategic advantage by enabling real-time object recognition, aiding in the precise allocation of resources, and ultimately enhancing the safety and effectiveness of military operations.
This article aims to explore how we train a sophisticated model called YOLO to recognize military objects. We are going to talk about the methods we use, the difficulties we face, and what could happen as a result. The goal is to show how important AI and computer vision are in changing how military technology works, which has big effects on keeping things safe and defending our countries.
Various methods are used to search for objects in the image when applying computer vision methods.
There are three main methods of object
recognition in the image:
• Contour analysis [
One of the methods for recognizing objects from a video stream is using pattern search methods. This method has information about what the required object looks like, what kind of background it can have, how certain contours of the object look, and at what positions they can be, immediately considering the possible location of the object detection. It allows us to achieve a high quality of recognition and has a good speed. However, when the video camera captures several objects that are similar to each other, different patterns are satisfied and recognition decreases. Therefore, apply a family of models to estimate functions.
Google Collaboratory [
A workspace will be organized before the project is developed. To identify enemy targets, DataSet padding will be performed for the corresponding data set. This process will include searching for images and videos of the above-mentioned objects and marking the corresponding objects. The data set will consist of tens of thousands of unique images—approximately equally for each class.
One of the methods used to determine moving objects is contour analysis [16], which is a method of describing, storing, recognizing, comparing, and searching for graphic images (objects) based on their contours. The contour completely defines the shape of the image and contains all the necessary information for recognizing images by their shape. This approach allows you not to consider the internal points of the image and thereby significantly reduce the amount of information that is transformed. Contour—a curve that describes the boundary of the object in the image. Therefore, it is possible to consider the contours of objects, which reduces the complexity of algorithms. The main advantage of the contour analysis is the invariance concerning the rotation, scale, and shift of the contour in the image.
It is well suited for searching for an object of a given shape. As a result, it is often possible to ensure the system works in real-time. However, there are significant disadvantages of this method. There are breaks in the outline in places. Thus, the contour analysis has a rather weak resistance to interference, and any violation of the integrity of the contour or poor visibility of the object leads to either the impossibility of detection or false positives. Simplicity and speed of contour analysis allow you to successfully apply this approach, provided there is a well-defined object on a contrasting background.
The template comparison method [17] is performed according to the criterion of the minimum (maximum) of some function comparing the object and its template. This method is one of the simplest. The input parameters of the method are the image on which the template must be searched, the image of the object that must be found on the tested image, and the size of the template must be smaller than the size of the tested image. The purpose of the algorithm is to find a section on the tested image that matches the template. Searching for a template is done by sequentially moving it by one pixel across the image, and evaluating the similarity of each new area to the template. Based on the results of the check, the section with the highest coincidence ratio is selected. In essence, this is the percentage of overlap between the image area and the template. The described method of matching with templates is simple, but there is a certain complexity in the process of creating templates, that is, in learning.
Template matching does not allow you to say whether the original object was found because it is a probabilistic characteristic that depends on the scale, viewing angles, rotations of the image, and the presence of physical obstacles [18]. There are also possible false positives of the algorithm when the searched object is not there, but there are some general details in the pattern and area on the tested image. Of course, this situation can be avoided by checking the match factor value (so that it is not less than some threshold), but this will not always work properly.
Often algorithms use key points [19] (feature points) of the image for their work. Key points are understood as some areas of the picture that are distinctive for this image. There are a large number of methods for detecting such “special points”, all of them differ in the speed of operation, the number of selected points, as well as resistance to image transformations: rotation, changes in viewing angles, changes in scale.
Three components are used to find key points in images and their subsequent comparison: • Feature detector[20]—searches for key points on the image. • Descriptor (descriptor extractor)— produces a description of the found key points, evaluating their positions through the description of the surrounding areas. • Matcher (matcher)—builds correspondences between two sets of image points.
Unlike template matching and contour analysis, algorithms for finding key points are more resistant to obstacles, and transformations and allow finding objects even in the presence of physical obstacles.
To achieve the highest possible level of tracking of an object (marker), it must have a significant number of unique (stable) key points, which the augmented reality library can quickly highlight in the video stream and compare with the existing template set. For this, it is necessary to use the fastest possible detector and descriptor, as well as to develop an algorithm that could confidently say that the object has been found.
This algorithm allows you to recognize images at different angles, at different distances from the camera, under different lighting, and when the image is partially overlapped [21].
object detection purposes. YOLOv8 is available in four primary versions: small (s), medium (m), large (l), and extra large (x), with each version providing increasing levels of accuracy.
Additionally, each variant requires a distinct amount of time for the training process.
The objective of the graph is to create a highly efficient object detector model, with performance indicated on the Y-axis and inference time on the X-axis. Initial findings indicate that YOLOv8 performs exceptionally well in achieving this goal compared to other cutting-edge techniques.
Examining the chart, it’s evident that all versions of YOLOv8 exhibit quicker training times than EfficientDet. Specifically, the most accurate YOLOv8 model, YOLOv8x, demonstrates the capability to process images at a significantly faster rate while maintaining a comparable level of accuracy compared to YOLO performs the prediction of both class the EfficientDet D4 [23] model. probabilities and bounding box coordinates. The architecture of YOLO involves dividing the YOLO assumes multiple bounding boxes for input image predicting bounding boxes, into a each grid cell. During training, the aim is to grid and and class probabilities. ensure that each bounding box predictor is
Here’s a high-level overview of the YOLO “responsible” for predicting the object based architecture: on the prediction with the highest value of
Input Layer: YOLO takes an input image, sample overlap with the correct information. typically of fixed size. To improve object detection accuracy,
Dividing into Grid: The image is divided YOLO uses the Non-Maximum Suppression into an S×S grid. Each grid cell is responsible (NMS) method [25]. This technique allows you for predicting bounding boxes and class to identify and remove redundant or incorrect probabilities. bounding boxes that may be created for a
Bounding Box Prediction: Each grid cell single object. As a result of applying NMS, only predicts multiple bounding boxes (usually B one bounding box is selected for each object in bounding boxes). These bounding boxes the image, which improves the quality and include information about the coordinates (x, efficiency of the detection process. y) of the bounding box, width (w), height (h), and confidence scores. 3.1. The Difference Between YOLO and
Class Prediction: For each bounding box, Other Deep Learning Algorithms for the model predicts class probabilities for Object Detection different object categories. This is done using softmax activation. The main difference between YOLO algorithms
Confidence Score: Each bounding box used for object detection is that it recognizes prediction is associated with a confidence objects quickly in real-time. The principle of score, indicating how likely it is that the operation of YOLO involves entering the entire bounding box contains an object. This score image at once, which passes through the takes into account both the probability of convolutional neural network only once [26]. object presence and the accuracy of the The performance of the YOLO algorithm is bounding box. evaluated using the COCO dataset [27] (Common
Final Prediction: The final predictions are Objects in Context). The COCO dataset is a large obtained by combining the bounding box dataset consisting of 80 feature classes. YOLOv1, coordinates, class probabilities, and confidence the baseline version, can recognize 24 object scores. Non-maximum suppression is then classes and has a 21.6% mAP (average accuracy) applied to filter out redundant and low- measured using the COCO dataset. YOLOv2 can confidence predictions. recognize 90 feature classes with a COCO dataset
Output: The final output is a set of bounding mAP of 30.2%. YOLOv3 can recognize 1000 boxes, each associated with a class label and a feature classes with an mAP of 57.9%. YOLOv4 confidence score. has a better performance compared to YOLOv3
Loss Function: YOLO uses a multi-part loss but can recognize 80 feature classes with mAP of function that includes terms for object 60.0%. YOLOv8 achieves significant presence/absence, bounding box coordinates, performance improvements over YOLOv4 and and class probabilities. This allows the model achieves an mAP of 83.5% on the COCO dataset. to be trained for accurate detection. YOLOv6 achieved 84.4% of the COCO mAP
The YOLO algorithm takes an image as input dataset. YOLOv7 achieved a mAP of 85.4%. and uses a deep convolutional neural network YOLOv8 provides additional performance to detect objects in the image. improvements over YOLOv7 and achieves an
The first 20 convolutional layers of the mAP of 86.4%. model are pre-trained using ImageNet [24], However, the YOLO algorithm has its using a temporal mean pooling layer and a fully drawbacks. Spatial limitations allow only connected layer. After that, this trained model 8Bbox to be projected per grid cell, making it is transformed to perform the object detection difficult to distinguish objects that are close task, to the trained network improves its together. Multiple samples are used and a lack performance. The last fully connected layer of of detail is often apparent. The third problem is imprecise localization, and finally, since Bbox training is performed based on data, it is difficult for the algorithm to detect a test dataset that does not exist in the training data.
However, the most difficult aspect of deep learning is the preparation of training data, and the data applicable to each application domain is very limited.
Despite all the limitations, the YOLO model still has a significantly higher processing speed than other models and continues to be widely used.
computing systems similar to the corresponding systems of the human brain. They consist of artificial neurons that receive input signals from a certain number of connections to the neurons of the previous layer or the input of the network. The scheme of a simple neural network with one hidden layer is shown in Fig. 1. The process of learning a neural network consists of adjusting the parameters of the network to obtain better results. For most problems in neural networks, tutoring is used, when the system is fed data with a ready result to gradually train the network with real data, where the correct answer will not be known in advance. Improving the results and accuracy of the network is done by minimizing a certain error, the difference between the predicted and actual outputs.
Accuracy measures how well the model predicts the correct outcome compared to the actual outcome. It is calculated by dividing the number of correct predictions by the total number of predictions. Although accuracy is a useful metric, it may not be sufficient to evaluate the performance of complex AI models. Therefore, it is important to consider other metrics such as precision, repeatability, and F1-score [28].
Classification errors—confusion matrix [29] (error matrix). If there are two classes and an algorithm that predicts each object to one of the classes, then the classification error matrix will look like this: where ̌ is the answer of the algorithm on the object, and y is the true label of the class on this object.
Accuracy measures the ratio of correct positive predictions to all positive predictions made by the model. It shows how well the model will avoid false positives. Repeatability, on the other hand, measures the ratio of correct positive predictions to all actual positive cases. It shows how well the model will avoid false negatives.
In this way, classification errors are applied, which are: False Negative (FN) and False Positive (FP) [29].
= =
+ × 100%, × 100%.
+
The F1-score is a combination of precision and repeatability, providing a balanced estimate of model performanceP).
1 = 2×+× × 100%. (3)
Reference data sets are traditionally used for preliminary comparison of models. The COCO dataset uses a special mAP averaged accuracy metric .
1 = ∫ ( ) ,
0 = 1 ∑3=1 + . (1) (2) (4) (5)
We will use an artificial neural network to build a classifier.
A neural network based on a multilayer perceptron is a system of interconnected layers of neurons. Each neuron is characterized by an activation function that converts the neuron’s input signal into an output signal. Connections of neurons with other neurons are characterized by connection coefficients—weights. An important factor in learning a neural network is the type of input data. To achieve the best results, it is necessary to preliminarily display the data using centering and scaling operations (1):
of operations for calculating the output signal of the network and subsequently changing the weights of the connections. As an algorithm for weight adjustment in MLP-based networks, the error backpropagation algorithm is usually used. It refers to methods of learning with a teacher, so it requires that target values be set in the training examples. This algorithm belongs to the class of gradient algorithms, that is, the changes in the weights of connections are made in the direction of minimizing the gradient of the error. The prediction error during training is equal to the difference between the signal at the network output and the output reference value corresponding to the input data (2).
= ( − ). (7)
out until the average value of the error during one training epoch decreases. Further training usually leads to a deterioration in the analytical capabilities of the neural network.
The network was trained using the error backpropagation algorithm and the gradient descent algorithm. The basis of the idea of the algorithm is the use of the initial error of the neural network (7) to calculate the correction values of the neuron weights in its hidden layers:
E = 1 2 ∑ (y−y′) k 2 i = 1, (8) where k is the number of output neurons of the network, y is the target value; and y’ is the actual output value.
This algorithm is iterative, it uses step-bystep learning. It works as follows: one test example is given to the learning input. Then the weight values are adjusted. At each iteration, forward and reverse passes of the network take place. On a forward input, the vector propagates from the inputs to the outputs of the network. In this way, a certain output vector is formed, which corresponds to the current (actual) state of the scales. The error of the neural network is calculated as the difference between the actual and target values.
On the return pass, this error is propagated from the output of the network to its inputs, and the neuron weights are corrected according to formula (4): Δwji(n) = −η ∂Eav ∂wij, (9) where wji is the weight of the ith connection of the jth neuron; η is a learning speed parameter that allows you to additionally control the size of the correction step; Δwji for more accurate adjustment to the minimum error and is selected experimentally in the learning process (changes in the interval from 0 to 1).
to optimization algorithms and is used to adjust the parameters of the machine learning model. The gradient is usually considered the sum of the gradients caused by each training element. The parameter vector changes in the direction of the anti-gradient with a given step. Therefore, standard gradient descent requires one pass over the training data before it can change the parameters. In stochastic (or “operational”) gradient descent, the value of the gradient is approximated by the gradient of the cost function calculated on only one training element. The parameters then change in proportion to the approximate gradient. Thus, the parameters of the model change after each training object. For large datasets, stochastic gradient descent can provide a significant speed advantage over standard gradient descent.
We will use the Python language to build the classifier. One of the main reasons why Python is used for machine learning is that it has many frameworks that simplify the process of writing code and reduce development time.
frame [30], and we want to save this video. We do pre-processing of images [31].
We create a VideoWriter object. Specify the name of the output file (output.avi). Then we specify the FourCC code and transfer the number of frames per second (fps) and the frame size. And the last one is Color. If True, the encoder expects a color frame, otherwise it works with a grayscale frame. FourCC is a 4byte code used to identify the video codec. XVID codec is better. MJPG creates large-size videos. X264 gives a small video size). On Windows: DIVX.
As part of the research, a proprietary dataset was utilized, comprising over a thousand images of tanks from various models and perspectives. These images were sourced from electronic books and websites, rendering them diverse and realistic [32]. Accordingly, it was necessary to manually add labels to the images for subsequent model training. This process ensured proper classification and recognition of tanks in the images.
The decision was made to employ the Roboflow platform for efficient uploading and processing of images.
the task but also accurately assigned labels to each object in the images, providing the model with sufficient information for tank recognition and classification.
and the Roboflow platform hold significant importance in advancing the tank recognition model, particularly in enhancing its accuracy and efficiency during training.
As a result of training, the model
demonstrated impressive accuracy, achieving
a high level of correct classification of objects
in the images [33]. The graphs attached to the
article illustrate the model's high stability and
effectiveness during training, confirming its
ability to confidently recognize tanks in
various scenarios and conditions [34].
Furthermore, the trained model exhibited
minimal losses during training, indicating high
efficiency and optimal adaptation to the
training data. The training and loss graphs
depict a stable learning process and the
absence of significant fluctuations in losses
during model optimization [
Overall, the obtained results affirm the high potential and efficiency of the developed model for tank recognition, making it a significant contribution to the field of military applications and security objects.
The model was evaluated using diverse images of
tanks, encompassing various models sourced
from different outlets [
We also conducted testing using a video stream to assess the model’s real-time tank recognition capabilities. In this mode, the model proved to be quite effective, adapting quickly to changes in the images and confidently recognizing tanks in different scenarios. Despite the overall success, it is worth noting that the model struggled to recognize tanks in low-light conditions, such as during nighttime without thermal imaging illumination or in snowy conditions. This aspect should be considered as a potential area for further improvements and optimizations to the model.
stages aimed at creating and refining a tank recognition model using YOLO (You Only Look Once). We generated our dataset, consisting of diverse tank images from various sources. This dataset provided crucial material for training and recognizing different tank models. [5] Utilizing the created dataset, we successfully trained the YOLO model to recognize tanks.
The training methodology involved annotation and systematic training to achieve high accuracy. We assessed the model’s effectiveness across various media, including photos and videos. The model demonstrated a high level of tank recognition in real-time and [6] images. The testing results revealed impressive accuracy and efficiency in recognizing tanks under different conditions and perspectives. Upon analysis, the next step could involve refining the model for tank recognition in low-light conditions, such as nighttime without thermal illumination, or during adverse weather conditions like [7] snowfall. In conclusion, the developed model holds significant potential in military applications and security domains. Further enhancements could make it even more effective in real-world conditions. [8] (2024) 17–20. doi: 10.55524/ijircst.2024. Recognition with YOLO: Advanced 12.1.4. Techniques in Robotic Object Detection, [14] U. Utkarsh, et al., Automated Translation Cognitive Robotics (2024). doi: and Accelerated Solving of Differential 10.1016/j.cogr.2024.01.001. Equations on Multiple GPU Platforms, [26] Y. Wang, et al., GT-YOLO: Nearshore Comput. Methods Appl. Mech. Eng. 419 Infrared Ship Detection Based on (2024) 116591. doi: 10.48550/arXiv. Infrared Images, J. Marine Sci. Eng. 12(2) 2304.06835. (2024) 213. doi: 10.3390/jmse12020213. [15] P. Feng, et al., Co-Continuous Structure [27] K. Gupta, A. Asthana, Reducing the SideEnhanced Magnetic Responsive Shape Effects of Oscillations in Training of Memory PLLA/TPU Blend Fabricated by Quantized YOLO Networks, IEEE/CVF 4D Printing, Virtual Phys. Prototyp. Winter Conference on Applications of 19(1) (2024) e2290186. doi: 10.1080/ Computer Vision (2024) 2452–2461. 17452759.2023.2290186. [28] P. Giudici, M. Centurelli, S. Turchetta, [16] J. Malik, et al., Contour and Texture Artificial Intelligence risk measurement, Analysis for Image Segmentation, Int. J. Expert Systems with Applications 235 Comput. Vision 43 (2001) 7–27. doi: (2024) 121220.
10.1023/A:1011174803800. [29] S. Shinde, et al., Artificial Intelligence [17] N. Hashemi, et al., Template Matching Approach for Terror Attacks Prediction Advances and Applications in Image Through Machine Learning, Multidiscip.
Analysis, arXiv preprint (2016). Sci. J. 6(1) (2024) 2024011–2024011. [18] G. Cox, Template Matching and Measures [30] M. Nazarkevych, et al., Evaluation of the of Match in Image Proce-Ssing, University Effectiveness of Different Image of Cape Town (1995). Skeletonization Methods in Biometric [19] D. Lowe, Distinctive Image Features from Security Systems, Int. J. Sens. Wirel.
Scale-Invariant Keypoints, Int. J. Comput. Commun. Control 11(5) (2021) 542–552.
Vision 60 (2004) 91–110. doi: 10.2174/221032791066620121015 [20] D. Mukherjee, M. Wu, G. Wang, A 1809.
Comparative Experimental Study of [31] M. Nazarkevych, et al., The Ateb-Gabor Image Feature Detectors and Filter for Fingerprinting, International Descriptors, Mach. Vision Appl. 26 Conference on Computer Science and (2015) 443–466. doi: 10.1007/s00138- Information Technology (2019) 247– 015-0679-9. 255. doi: 10.1007/978-3-030-33695[21] L. Liu, et al., CLFR-Det: Cross-Level 0_18.
Feature Refinement Detector for Tiny- [32] M. Nazarkevych, et al., Data Protection Ship Detection in SAR Images, Based on Encryption Using AtebKnowledge-Based Syst. 284 (2024). doi: Functions, 9th International Scientific 10.1016/j.knosys.2023. 111284. and Technical Conference Computer [22] Q. Liu, et al., YOLOv8-CB: Dense Sciences and Information Technologies Pedestrian Detection Algorithm Based (2016) 30–32. on In-Vehicle Camera, Electronics 13(1) [33] M. Medykovskyy, et al., Methods of (2024) 236. doi: 10.3390/electronics Protection Document Formed from 13010236. Latent Element Located by Fractals, Xth [23] F. Pan, et al., Zero-shot Building Attribute International Scientific and Technical Extraction from Large-Scale Vision and Conference Computer Sciences and Language Models, IEEE/CVF Winter Information Technologies (2015) 70–72. Conference on Applications of Computer [34] V. Sheketa, et al., Formal Methods for Vision (2024) 8647–8656. Solving Technological Problems in the [24] H. Li, C. Wang, Y. Liu, YOLO-FDD: Infocommunications Routines of Efficient Defect Detection Network of Intelligent Decisions Making for Drilling Aircraft Skin Fastener, Signal, Image and Control, IEEE International ScientificVideo Processing (2024) 1–15. Practical Conference Problems of [25] S. Koga, et al., Optimizing Food Sample Infocommunications, Science and Handling and Placement Pattern Technology (2019) 29–34.