Feature extraction network of water scenes based on improved SuperPoint

Since water scenes are highly susceptible to environmental factors, in the actual process of collecting and storing pictures, it is easy to cause noise pollution of image samples. This requires that the feature extraction algorithm can reduce the impact of noise on the sample data when dealing with data containing noise, i.e.That is to say, the algorithm model is required to have high robustness. The high-speed corner detection algorithm FAST, scale-invariant feature transform SIFT, etc. are traditional feature point detection methods, but the advantages in terms of computational speed and robustness are mixed. In this paper, we focus on the SuperPoint network, which has better robustness, and modify the network accodrding to the requirements of real-time and accuracy. To address the problems of gradient vanishing and gradient explosion, a residual connection structure is added between each convolutional layer and activation function. Meanwhile, to ensure the convergence speed of the model, a normalisation layer is added between the convolutional layers and the activation function. Finally, in order to improve the representation and generalisation ability of the model, the SE-Net channel attention mechanism module is added after the residual connection structure. Some of the Seaship ships, navigation beacons and other datasets are transformed into feature point datasets in the training dataset and strengthen its feature extraction capability for water scenes.Experimental analysis is conducted based on the water scenario, and the experimental results show that the detection and matching effects of feature points are improved under the guarantee of a slight increase in the computational speed, the detection effect of the number of feature points is improved by about 15.4%, the matching effect of feature points is improved by about 9.7%, the nearest-neighbour accuracy of the NN mAp is improved by about 9%, the repeatability of the Rep. is improved by about 11.4%, and the average positioning error is reduced by about 2.1%.