Mine Detection Using CNN+BiLSTM+Attention Based on B-Scan Signals

In modern conditions, the problem of mine detection remains one of the most urgent due to the serious threat to the life and health of people in contaminated areas. This paper presents an approach to mine detection using a hybrid neural network CNN+BiLSTM+Attention, which analyzes B-scan signals received from ground-penetrating radar systems. To improve the quality of training with a limited amount of data, image augmentation was used, which includes shifting, reflecting, scaling, and adding noise. The initial layers of the architecture use convolutional operations to extract local spatial features, after which a bidirectional recurrent layer BiLSTM is used, which allows the model to learn dependencies within each B-scan, taking into account the context in both directions. An attention mechanism is additionally integrated to focus on the most informative fragments of the signal. The final layers of the model are dense layers with a sigmoid activation function for mine detection. The results of the computer experiment demonstrated the high efficiency of the model: the classification accuracy exceeds 99%, and the values of Precision, Recall, F1-score and AUC indicate the reliability of the approach. Visualization of the results (loss and accuracy graphs, confusion matrix, metric histogram) confirms stable learning without overtraining. The proposed architecture is promising for application in remote mine monitoring tasks and can be used as a basis for further research in the field of military security.