Trainable visual navigation systems based on deep learning demonstrate potential for robustness of onboard camera parameters and challenging environment. However, a deep model requires substantial computational resources and large labelled training sets for successful training. Implementation of the autonomous navigation and training-based fast adaptation to the new environment for a compact drone is a complicated task. The article describes an original model and training algorithms adapted to the limited volume of labelled training set and constrained computational resource. This model consists of a convolutional neural network for visual feature extraction, extreme-learning machine for estimating the position displacement and boosted information-extreme classifier for obstacle prediction. To perform unsupervised training of the convolution filters with a growing sparse-coding neural gas algorithm, supervised learning algorithms to construct the decision rules with simulated annealing search algorithm used for finetuning are proposed. The use of complex criterion for parameter optimization of the feature extractor model is considered. The resulting approach performs better trajectory reconstruction than the well-known ORB-SLAM. In particular, for sequence 7 from the KITTI dataset, the translation error is reduced by nearly 65.6% under the frame rate 10 frame per second. Besides, testing on the independent TUM sequence shot outdoors produces a translation error not exceeding 6% and a rotation error not exceeding 3.68 degrees per 100 m. Testing was carried out on the Raspberry Pi 3+ single-board computer.

Neural network object detectors are increasingly being used for aerial video analysis, with a growing demand for onboard processing on UAVs and other limited resources. However, the vulnerability of neural networks to adversarial noise, out-of-distribution data, and fault injections reduces the functionality and reliability of these solutions. The development of detector models and training methods that simultaneously ensure computational efficiency and robustness against disturbances is an urgent scientific task. The research subjects. The model and method for ensuring the robustness of resource-constrained neural network systems for object detection in aerial video surveillance. Objective. Development of a model and method to ensure the robustness of object detectors for aerial image analysis. Methods. Combination of ideas and methods for dynamic neural networks, and methods for robustness and resilience optimization for neural networks. Results. The detector model with a ViT-B/16 backbone modified with gate units for dynamic inference was developed. The model was trained on the VEDAI dataset and meta-trained on the results of adaptation to different types of disturbances. The model with different training methods was tested for robustness against random bit-flip injection where the proportion of the modified weights is determined at a fault rate of 0.1. In addition, the model with different training methods were tested for robustness against a black-box Adversarial Attack with a perturbation level of 3/255 according to the L¥ norm. Conclusions. The object detection model for aerial images with dynamic inference and optimized robustness is developed for the first time. The model includes a transformer-based backbone, gate units, and simplified feature pyramid network with a RetinaNet detection head. Gate units are trained to deactivate transformer encoders that are irrelevant to the input data and disturbances. The proposed model reduces FLOPs by more than 22% without loss of mean Average Precision (mAP) by deactivating some encoders. The detector training method was developed for the first time, combined the RetinaNet loss function with the gate unit loss function and applied meta-learning to the results of adaptation to various types of synthetic disturbances. The analysis of the experimental results demonstrates that the proposed method provides an 11.7 % increase in mAP during testing under fault injection conditions and a 15.1 % increase in mAP during adversarial attack testing.

The efficient model and learning algorithm of the small object detection system for compact aerial vehicle under conditions of restricted computing resources and the limited volume of the labeled learning set are developed. The four-stage learning algorithm of the object detector is proposed. At the first stage, selecting the type of deep convolutional neural network and the number of low-level layers that is pretrained on the ImageNet dataset for reusing takes place. The second stage involves unsupervised learning of high-level convolutional sparse coding layers using the modification of growing neural gas to automatically determine the required number of neurons and provide optimal distributions of the neurons over the data. Its application makes it possible to utilize the unlabeled learning datasets for the adaptation of the high-level feature description to the domain application area. At the third stage, the output feature map is formed by concatenation of feature maps from the different level of the deep convolutional neural network. At that, there is a reduction of output feature map using principal component analysis and followed by the building of decision rules. In order to perform the classification analysis of output, feature map is proposed to use information-extreme classifier learning on principles of boosting. Besides that, the orthogonal incremental extreme learning machine is used to build the regression model for the predict bounding box of the detected small object. The last stage involves fine-tuning of high-level layers of deep network using simulated annealing metaheuristic algorithm in order to approximate the global optimum of the complex criterion of learning efficiency of detection model. As a result of the use of proposed approach has been achieved 96% correctly detection of objects on the images of the open test dataset which indicates the suitability of the model and learning algorithm for practical use. In this case, the size of the learning dataset that has been used to construct the model was 500 unlabeled and 200 labeled learning samples

A machine learningsemi-supervised method was developed for the classification analysis of defects on the surface of the sewer pipe based on CCTV video inspection images. The aim of the research is the process of defect detection on the surface of sewage pipes. The subject of the research is a machine learning method for the classification analysis of sewage pipe defects on video inspection images under conditions of a limited and unbalanced set of labeled training data. A five-stage algorithm for classifier training is proposed. In the first stage, contrast training occurs using the instance-prototype contrast loss function, where the normalized Euclidean distance is used to measure the similarity of the encoded samples. The second step considers two variants of regularized loss functions – a triplet NCA function and a contrast-center loss function. The regularizing component in the second stage of training is used to penalize the rounding error of the output feature vector to a discrete form and ensures that the principle of information bottlenecking is implemented. The next step is to calculate the binary code of each class to implement error-correcting codes, but considering the structure of the classes and the relationships between their features. The resulting prototype vector of each class is used as a label of image for training using the cross-entropy loss function. The last stage of training conducts an optimization of the parameters of the decision rules using the information criterion to consider the variance of the class distribution in Hamming binary space. A micro-averaged metric F1, which is calculated on test data, is used to compare learning outcomes at different stages and within different approaches. The results obtained on the Sewer-ML open dataset confirm the suitability of the training method for practical use, with an F1 metric value of 0.977. The proposed method provides a 9 % increase in the value of the micro-averaged F1 metric compared to the results obtained using the traditional method.

The subject of study in this article is neural network–based methods for aerial image matching, which are widely used in navigation, localization, and mapping tasks. A key challenge lies in the sensitivity of such methods to visual disturbances and scene novelty caused by shadows, illumination changes, and terrain variability, which limits their robustness in real-world conditions—particularly under constrained computational resources. This paper investigates an approach to enhancing the robustness and cross-domain generalization of computationally efficient aerial image matching models by combining adversarial procedural noise with a modified activation function. The goal is to develop a training methodology that simultaneously increases the resilience of models to perturbations and improves their transferability across different observation domains. The research objectives are as follows: (1) to analyze existing methods for improving the robustness of neural networks and assess their applicability to aerial image matching tasks; (2) to develop a training approach incorporating the synthesis of adversarial procedural noises (Perlin, Gabor, Worley) and the replacement of the standard ReLU with a hybrid activation function, LeakyReLU6, which constrains activation amplitudes and reduces sensitivity to local disturbances; (3) to conduct a comprehensive experimental evaluation of detector-based architectures (SuperPoint + LightGlue) and detector-free models (EfficientLoFTR) using the Aerial Image Matching Benchmark dataset; (4) to verify cross-domain generalization on the HPatches dataset; and (5) to perform an ablation study to isolate the contribution of each component. Results. The proposed methodology achieved over a 4.2% absolute improvement in AUC@1px matching accuracy on noisy test data for both classes of models. The ablation study revealed a synergistic effect from combining procedural noise with LeakyReLU6 — in particular, for the SuperPoint + LightGlue combination, improvements reached +3.0% AUC@1px and +2.7% AUC@3px, while for EfficientLoFTR, gains of +2.2% and +2.6% were observed, respectively. Additionally, testing on HPatches showed a 0.83% smaller performance drop compared to baseline training, confirming a higher level of cross-domain generalization. Conclusions. The proposed approach enhances the noise robustness and cross-domain generalization of feature-matching models and can be easily extended to various neural network architectures. Future work will focus on investigating the influence of procedural noise hyperparameters, applying meta-learning on corrupted data, and introducing architectural improvements to further strengthen resilience and robustness. Scientific novelty. The novelty of this work lies in the first integration of adversarial learning with procedural noise and a bounded activation function (LeakyReLU6, using the Straight-Through Estimator (STE) in the backward pass), which produced a synergistic effect that improved the robustness and generalization of aerial image matching models without a significant increase in computational cost.

Models and training methods for water-level classification analysis on the footage of sewage pipe inspections have been developed and investigated. The object of the research is the process of water-level recognition, considering the spatial and temporal context during the inspection of sewage pipes. The subject of the research is a model and machine learning method for water-level classification analysis on video sequences of pipe inspections under conditions of limited size and an unbalanced set of training data. A four-stage algorithm for training the classifier is proposed. At the first stage of training, training occurs with a softmax triplet loss function and a regularizing component to penalize the rounding error of the network output to a binary code. The next step is to define a binary code (reference vector) for each class according to the principles of error-correcting output codes, but considering the intraclass and interclass relations. The computed reference vector of each class is used as the target label of the sample for further training using the joint cross-entropy loss function. The last stage of machine learning involves optimizing the parameters of the decision rules based on the information criterion to account for the boundaries of deviation of the binary representation of the observations of each class from the corresponding reference vectors. As a classifier model, a combination of 2D convolutional feature extractor for each frame and temporal network to analyze inter-frame dependencies is considered. The different variants of the temporal network are compared. We consider a 1D regular convolutional network with dilated convolutions, 1D causal convolutional network with dilated convolutions, recurrent LSTM-network, recurrent GRU-network. The performance of the models is compared by the micro-averaged metric F1 computed on the test subset. The results obtained on the dataset from Ace Pipe Cleaning (Kansas City, USA) confirm the suitability of the model and training method for practical use, the obtained value of F1-metric is 0.88. The results of training by the proposed method were compared with the results obtained using the traditional method. It was shown that the proposed method provides a 9 % increase in the value of micro-averaged F1-measure.