Attention-based mechanism for SuperPoint feature point extraction in endoscopy

Routine endoscopes have been widely used in medical diagnosis. Three-dimensional (3D) modelling reconstruction of endoscopic images has become the development direction of future medical domain. Local feature extraction and matching is a key step for 3D modelling reconstruction. Handcrafted local features such as SIFT, SURF, ORB, are still a predominant tool for such tasks. Due to the special environment of endoscopes, there are generally weak textures and large lighting changes, which make traditional feature point extraction algorithms unable to extract feature points well. We explore the potential of the self-supervised method SuperPoint. Many existing works have shown the benefits of enhancing spatial encoding. We propose a new architecture unit, in which the SE attention mechanism module is proposed, which can explicitly model the interdependence between convolutional feature channels to improve the network's representation ability. The experimental results show that this multi-scale channel attention feature point extraction algorithm based on SuperPoint has better result and achieves higher matching quality than handcrafted local features and original algorithm in endoscopic images.