Deng entropy and LSTM neural network to classify built cultural heritage with severe damage

This paper proposes an automatic approach for the 3D recognition of the structural state of uniquely built cultural heritage with severe and very severe damage. For this purpose, a 3D reconstruction (point cloud) mapped to voxels of built cultural heritage is used. Then, Deng entropy is computed for the resulting voxel map, and the corresponding information dimension is obtained by applying the box-covering method. The entropy and information dimension data sequences are combined into one sequence, called the Deng sequence. This is in order to quantify the occupied space in voxel map of built cultural heritage. Deng sequence is used as the input for a bidirectional Long Short-Term Memory neural network (bLSTMnn) to automatically classify the structural state of built cultural heritage. The main consideration is that few models of built cultural heritage are available due to their uniqueness. Hence, a small dataset created by authors using Computer-Aided Design tools is used for the training of the bLSTMnn. A cross-method is applied to validate the bLSTMnn. The proposed approach is successfully applied to recognize the current structural state of a unique Mexican cultural heritage building with very severe damage.