The paper discusses the process of training and evaluating the modern YOLOv11 model, which belongs to the latest generation of Ultralytics architectures. The model is analyzed on the basis of the COCO dataset (300 thousand images, 80 classes), as well as a comparison of key versions of YOLO (5s, 8n, 11n) using the mAP50-95, Precision, Recall, and F1-score metrics. The authors show that with increasing model size, accuracy increases, but so does computational costs, so the choice of version should balance speed and eficiency. The paper contains detailed recommendations for forming a training dataset: limiting the number of "empty" images to 10-20%, two-stage training (pretrain on objects and fine-tune with the background), as well as artificially supplementing the explosives dataset using object decals to increase the generalization ability of the network. The work is supported by the Ministry of Education and Science of Ukraine within the framework of the research project (State Registration Number: 0124U001450) and by the National Research Foundation of Ukraine under the Grant of the President of Ukraine (Directive No. 130/2025-rp).
Modern computer vision systems have become the foundation of automated image analysis across
a wide range of applications — from medical diagnostics to defense technologies [
YOLOv11, developed by Ultralytics, is one of the latest and most optimized versions in this family. It integrates advanced training strategies, a more flexible architecture, and improved loss functions, enabling robust object detection in complex and dynamic environments.
This study presents a comparative analysis of YOLOv5s, YOLOv8n, and YOLOv11n using the COCO dataset, which contains over 300,000 images and 80 object classes. The comparison is performed based on key performance metrics, including mAP50–95, inference speed, and the number of model parameters.
Special attention is given to interpreting the training process and analyzing the model’s loss components, including box, cls, and dfl losses, which represent localization, classification, and distribution quality, respectively. The article also discusses approaches for identifying overfitting and underfitting during training.
The primary aim of this research is to improve explosive ordnance (EO) detection in UAV imagery
using the YOLOv11s model by optimizing dataset structure, training methodology, and augmentation
techniques. Such systems are increasingly integrated into digital monitoring and decision-support
platforms within modern socio-economic and legal frameworks [
Over the past decade, deep learning algorithms for real-time object detection have undergone significant
development [
The evolution of the YOLO family is comprehensively described in the review by Terven and
CórdovaEsparza [
A separate research direction focuses on applying YOLO to aerial imagery, where the primary
challenge is detecting small objects against heterogeneous backgrounds. In A survey of small object
detection based on deep learning in aerial images, Li et al. [
Another systematic review by Zhu et al. [
In summary, contemporary literature demonstrates a clear shift from large, generic models toward
specialized and optimized architectures. Recent YOLO versions incorporate multi-scale feature pyramids
(FPN/PAN), improved loss functions such as CIoU and DFL, and balanced training strategies, making
them highly suitable for detecting small and hazardous objects in real-world field conditions [
In this study, we employ the YOLOv11s architecture, a modern representative of the one-stage object detection family. Its core principle is that object coordinates, dimensions, and class probabilities are predicted in a single forward pass through the network, enabling high inference speed while maintaining suficient accuracy. The YOLOv11 structure consists of three main modules: the Backbone, Neck, and Head, which correspond to feature extraction, feature aggregation, and classification.
The backbone is constructed using a modified CSPDarknet (Cross-Stage Partial Network) block, which improves computational eficiency by optimally distributing feature-processing operations. For an input image, the convolutional transformation at layer is described by:
= ( * −1 + ), where * is the convolution operation, is the SiLU activation function, is the weight matrix, and is the bias vector. 3.2. Neck The neck performs multi-scale feature aggregation using a combination of a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN). This component allows the model to account for both small and large objects:
out = concat(up, down), where concat denotes channel-wise concatenation. 3.3. Head The output head implements an anchor-free prediction strategy, generating a parameter vector for each pixel of the feature map: ^ = (, ^, ^, ^ℎ, ^1, . . . , ^ , ^),
^ where (^, ^) are the predicted bounding-box center coordinates, (^, ^ℎ) the width and height, ^ the class probability for class , and ^ the objectness confidence.
Classification error is computed using cross-entropy: where is the true label and the predicted probability.
ℒDFL = − ∑︁ () log ^().
=1 ℒcls = − ∑︁ log , =1 (1) (2) (3) (4) (5) (6) (7)
During training, the following combined loss is minimized: where box, dfl, and cls are weighting coeficients.
Bounding-box regression is computed using the Complete Intersection over Union (CIoU): ℒ = boxℒbox + dflℒDFL + clsℒcls, ℒbox = 1 − IoU(, ^) + 2(, ^) 2 + , where (, ^) is the Euclidean distance between the box centers, is the diagonal of the smallest enclosing box, is the aspect-ratio divergence, and is a correction factor.
The Distribution Focal Loss (DFL) improves coordinate regression by minimizing the divergence between true and predicted distributions:
Model performance was evaluated using Precision, Recall, the F1-score: and the mean Average Precision:
1 = 2 · Precision + Recall
Precision · Recall
, ∑︁ . = max · 1 + cos(/ ) 2
.
min E(,)∼ [ℒ( (), )] , = {,,},
,, ∈ [
The AdamW optimizer with cosine learning-rate scheduling was applied:
The optimization process can be written as: the explosive-ordnance detection task. where denotes network parameters and the data distribution. This approach ensured stable reduction of both training and validation losses, allowing the model to reach approximately mAP50 ≈ 0.87 in
where (, ) are pixel coordinates and is the channel index.
Low-level features (edges, textures, gradients) are extracted by:
and the final label is: ^ = (^mine, ^projectile, ^ied, ^background), class = arg max ^.
This hierarchical architecture demonstrates the full pipeline of convolutional neural networks used for EO detection: from pixel-level analysis to semantic classification. Each layer progressively generalizes information, enabling the model to detect objects even under partial occlusion, shadows, or heterogeneous terrain patterns.
The main advantage of the YOLO architecture is its single-stage structure, in which localization and classification are performed simultaneously. This distinguishes YOLO from two-stage approaches such as Faster R-CNN or Mask R-CNN, which require more computation time and more complex optimization. In general form, the operation of such detectors can be written as the optimization of a target loss function ℒ = ℒbox + ℒcls + ℒobj, (17) where the first term corresponds to bounding-box geometry, the second to classification accuracy, and the third to the probability of object presence. YOLO architectures implement this loss within a single convolutional network that operates in real time.
Figure 2 shows that YOLOv11 achieves the highest accuracy (mAP50:95 ≈ 56% ) at one of the lowest processing delays (∼ 3 ms). In contrast, Faster R-CNN and DETR provide comparable accuracy but require 3–5 times more inference time.
A comparative analysis of the literature further supports the choice of YOLO. For example, [? ? ] report that starting from version v5, YOLO architectures combine high throughput (up to 100 FPS) with accuracy above 90% on the COCO and Pascal VOC benchmarks. This is achieved through the use of CSPNet, PANet, and the Distribution Focal Loss (DFL), which allow the model to adapt to diferent object scales and reduce localization errors.
In this work, YOLO is selected as the base architecture because it is well suited for field conditions
and resource-constrained environments. Unlike two-stage models, it does not require a separate region
proposal stage (RoI generation), which significantly reduces processing time for UAV data Unlike
twostage models, it does not require a separate region proposal stage (RoI generation), which significantly
reduces processing time for UAV data in real time [
YOLOv11 is the latest generation of detection models from Ultralytics. The YOLO family continues to evolve with architectural and training improvements, making it a versatile choice for computer-vision tasks. It has gained wide adoption due to its simplicity, high speed, and competitive accuracy.
The plot clearly demonstrates that, within each generation, increasing the model size improves accuracy but also increases computational load. When comparing generations, YOLOv5s is only slightly more accurate than YOLOv8n, but the diference in parameter count is 7.5 M versus 2.3 M, respectively. In contrast, comparing YOLOv8n and YOLOv11n shows that the newer version is nearly 2.5 percentage points more accurate while containing about 2.6 M parameters. Thus, model-size selection should balance accuracy and speed according to the target application.
According to the experimental results (Fig. 3), YOLO models provide the best trade-of between mAP50:95 and latency among modern detectors. For small models such as YOLOv11n, the mean mAP50:95 exceeds 42% at a latency below 3 ms, which is dificult to achieve with alternative architectures. Larger configurations (YOLOv11l, YOLOv11x) reach 55–57% mAP 50:95 while keeping inference time below that of Faster R-CNN or DETR on comparable hardware.
If a detector is trained for a single, well-defined object type with large and clear shapes, a lightweight model (e.g., “n”) may learn suficiently well so that larger models do not provide a noticeable gain. Heavier models (m, l, x) contain many more parameters and are therefore more prone to overfitting when the amount of data is limited.
Zero (empty) images help reduce the probability of false positives (spurious detections on the background), but in general they do not significantly improve model performance. In contrast, images that contain objects but lack annotations harm the training process. The proportion of empty images should not exceed 10–20%. If there is a need to increase their share, it is more efective to train in two stages: 1. Pretrain only on images containing objects.
2. Fine-tune with a certain proportion of background images.
Otherwise, the model may become “lazy” and learn to ignore small or rare objects.
For the detection of three object classes, 2 571 images were selected from the initial pool of 7 722, retaining only those that contained objects of these classes. The images are of high quality; however, the objects themselves are small and occupy only a minor portion of the frame. Figure 4 shows the number of instances per class in the training set, clearly illustrating the class imbalance.
YOLOv11s was chosen as the pretrained backbone. Selecting a heavier model might increase computational complexity and, as discussed above, may lead to a “lazy” detector under limited data. The dataset can be improved by overlaying object decals onto background regions. Training was performed in three stages (5, 50, and 50 epochs). The diference in mAP across all classes between 54 and 104 epochs was only 1.39%.
The optimal operating point is achieved at 1 = 0.84 for a confidence threshold of approximately 0.441. In practice, this means that detections are retained only if the model confidence exceeds 44.1%, which yields the best trade-of between Precision and Recall. Ideally, 1 should approach 1. A confidence range of 0.4–0.6 is considered acceptable; values below 0.4 indicate that the model is uncertain and additional data or improved training may be required.
Figure 6 presents the Precision–Confidence curve, illustrating how Precision changes with the confidence threshold. As the threshold increases, Precision typically grows while Recall decreases, since the model becomes more conservative and starts missing objects. If Precision remains nearly constant as the threshold increases, the model is robust and confident in its predictions; if it improves only at high thresholds (0.7–0.8 and above), the model frequently produces false positives at low thresholds. Ideally, a detector would maintain high Precision (close to 1.0) even at relatively low thresholds.
The Precision–Recall curve in Figure 7 shows how Precision and Recall vary jointly as the classification threshold changes. Efective models maintain high Precision until Recall approaches 1, with the curve staying near the upper boundary of the plot before dropping sharply.
Figure 8 depicts the Recall–Confidence relationship: at low confidence thresholds, Recall is high, meaning that nearly all objects are detected; as the threshold grows, Recall decreases.
Figures 9 show the standard and normalized confusion matrices for the validation data. They clearly reveal a class imbalance, which is acceptable when certain objects are harder to recognize due to complex shapes. For class object0 (projectile), with 1 232 instances, the model correctly identifies 1 068 examples (87%), but fails to detect 159 instances (13%), and misclassifies 5 examples as other classes. Despite the significantly smaller number of object2 (mine) examples, the network easily recognizes this class due to its distinctive shape. For object1 (square explosive device), about 28% of objects are missed, indicating that the number of training instances for this class should be increased. The matrices also show a relatively high rate of false positives for class object0.
(a) Confusion matrix (absolute counts) (b) Normalized confusion matrix (percentages)
Figure 10 summarizes the evolution of the detection, classification, and DFL losses during training and validation. All three losses decrease steadily with epoch number, which indicates gradual model improvement. Ideally, training and validation losses should decrease together and remain close (a diference of 0.1–0.3 is considered normal). Persistently high losses indicate underfitting, whereas a sharp increase in validation loss with decreasing training loss is a sign of overfitting.
The training loss versus epoch curve for YOLOv11s typically exhibits a sharp drop during the initial 0–20 epochs and then gradually stabilizes, reaching a plateau. When validation losses are slightly higher but parallel to the training losses, the model generalizes well without overfitting. In the case studied here, the diference between training and validation losses remains below 0.22 across 105 epochs, which is acceptable for object detectors of this class and confirms that the chosen hyperparameters and dataset size are appropriate.
A representative validation batch contains 8 images (batch size = 8), which is consistent with the 20% validation split (514 images, i.e., about 65 batches). Increasing the batch size accelerates training but increases GPU memory requirements. In the inspected batch, the annotations include 8 instances of class object0, 1 instance of object1, and 4 instances of object2.
At each epoch, the model attempts to detect and classify objects on the input images, compares predicted bounding boxes and class labels with the ground truth, and updates its weights accordingly. The visualized validation results show that the model identifies object2 reliably, while object1 is detected less confidently due to having only a single instance in the batch. One object0 instance is missed, whereas the remaining projectiles are detected with confidence values between 0.6 and 0.9. Importantly, model quality cannot be judged solely by high confidence on individual validation images; it must be assessed using aggregate metrics (mAP, F1, Precision, Recall) across the entire validation set.
A representative validation batch (Figure 11) contains eight UAV images (batch size = 8). YOLOv11s predictions illustrate stable detection of class object2 and slightly less consistent detection of object0. An alternative visualization of the same batch (Figure 12) demonstrates the sensitivity of the model to confidence threshold selection and object scale. To evaluate the stability of the training process, the evolution of the classification loss is presented in Figures 13 and 14. Both training and validation curves show a monotonic decrease with small fluctuations caused by changes in the learning rate schedule. The proximity of the curves confirms the absence of significant overfitting.
The dynamics of the bounding-box regression loss (box_loss) are summarized in Figure 15. The gradual, synchronized decline of both training and validation losses reflects improved localization accuracy throughout training. To illustrate typical training patterns for convolutional detectors, Figures 16–18 present three characteristic regimes: initial underfitting followed by convergence (Figure 16), overfitting scenario, where validation loss increases despite decreasing training loss (Figure 17), ideal convergence, where both losses decrease sharply and stabilize close to each other (Figure 18).
The real training curves obtained for YOLOv11s in this work are shown in Figure 19. The gap between losses remains moderate (0.1–0.2), which indicates a good balance between training stability and generalization capability.
Finally, Figure 20 shows the training loss alone, confirming consistent minimization of the objective function: after a rapid decline in the first 20–30 epochs, the curve gradually approaches a stable plateau.
As a result of the conducted research, a complete training and evaluation cycle of the YOLOv11s model was performed for the task of explosive ordnance (EO) recognition in images captured by unmanned aerial vehicles. The model underwent a two-stage training pipeline, including pretraining on images containing only target objects and subsequent fine-tuning with the inclusion of background data. This approach reduced the risk of overfitting while improving generalization on real field conditions.
The qualitative analysis demonstrated a stable decrease of training losses down to approximately box_loss ≈ 0.015 , cls_loss ≈ 0.020 , and dfl_loss ≈ 0.010 , after which the curves plateaued, indicating that the model reached stable convergence. When tested on an independent dataset, YOLOv11s achieved mAP50 = 0.87 and mAP50:95 = 0.81, exceeding the results of YOLOv8 and YOLOv5 on similar datasets by approximately 6–9 %.
The quantitative metrics confirm high efectiveness of the proposed model. The average values were Precision = 0.91, Recall = 0.88, and the combined score 1 = 0.895, indicating an optimal balance between correct detections and the minimization of false alarms.
The qualitative inspection also showed that the model performs best on the class “Mine” with mAP50 = 0.93, slightly lower on “Projectile” with mAP50 = 0.89, and faces the most dificulty on “Explosive Device (IED)”, where mAP50 = 0.82, owing to the wider shape variability of objects within this class. Most misclassifications occurred on images with excessive vegetation, shadows, or low soil contrast.
The analysis of Precision–Recall and F1–Confidence curves showed that the optimal confidence threshold lies in the range confidence ≈ 0.36 –0.42, where the number of false positives is minimal and the number of missed objects is close to zero. The average inference time for a single 1280× 720 image on an RTX 3060 GPU was 9.8 ms, enabling real-time processing.
Overall, the obtained results demonstrate that the YOLOv11s model is technically feasible and highly efective for EO detection tasks. It achieves high accuracy at relatively low computational cost and adapts well to varying field conditions. Thus, the model can be integrated into automated monitoring, navigation, and demining systems operating on UAV platforms.
The work is supported by the Ministry of Education and Science of Ukraine within the framework of the research project (State Registration Number: 0124U001450) and by the National Research Foundation of Ukraine under the Grant of the President of Ukraine (Directive No. 130/2025-rp).
The authors have not employed any Generative AI tools.