Search

Search Results

• CONVOLUTIONAL NEURAL NETWORK HYBRID ARCHITECTURE DEVELOPMENT USING SPECTRAL TRANSFORMATIONS

  B. V. Kostrov , S.I. Babaev , А.I. Efimov , V. Y. Tarasova

  2026-02-27

  Abstract ▼

  The hybrid convolutional neural network architecture with combining spectral and spatial layers, as well as new methods of subsampling (WalsPooling) and convolution (ConvWals) are proposed. The developed system is used to geographical proximity assess of images pair based on their visual similarity. A pair of different sensors obtained images visual similarity determination is complicated by different scales and sensor tilt angles shooting conditions. Based on the low-altitude image fragment, a search in the database of underlying surface images is performed. The search is performed in the surrounding area of a given route based on the vector of image features, which is formed on the last layer of the convolutional neural network. The system uses the Siamese architecture, since a pair of images must be submitted to the input. The relevance of this problem stems from the need to ensure UAV navigation in the absence or unreliability of a GPS signal. The approach to data set formation and its preprocessing is also considered. The database search is performed in the surrounding area of the route, which reduces computational costs. The experiments include an analysis of the applicability of the proposed layers (WalsPooling, ConvWals) and a comparison with traditional pooling and convolution methods. The paper also presents a linear approximation method with trainable parameters for reducing the dimensionality of the convolutional layer. The main advantage of the approach is its resistance to changes in the scale and angle of shooting due to a combination of spectral and spatial features. The results demonstrate the applicability of the method for UAV navigation in conditions of loss of GPS signal is lost or unreliable. The experiment demonstrated that using images reconstructed after spectral transformation yields the best neural network convergence and mean square error. The developed architecture demonstrates robustness to geometric and brightness distortions, and its quality metrics (Precision = 0.728, Recall = 0.800, F1 = 0.872) confirm the effectiveness of the approach for visual localization tasks based on images from a surface database.