Underwater target feature extraction method based on convolutional neural network (CNN)

Abstract

The invention provides an underwater target feature extraction method based on a convolutional neural network (CNN). 1, a sampling sequence of an original radiation noise signal is divided into 25 consecutive parts, and each part is set with 25 sampling points; 2, normalization and centralization processing are carried out on a sampling sample of the j-th segment of the data signal; 3, short-time Fourier transform is carried out to obtain a LoFAR graph; 4, a vector is assigned to an existing 3-dimensional tensor; 5, an obtained feature vector is input to a fully-connected layer for classification and calculation of error with label data, whether the loss error is below an error threshold is tested, if the loss error is below the error threshold, network training is stopped, and otherwise, step 6 is entered; and 6, a gradient descent method is used to carry out parameter adjustment layer by layer on the network from back to front, and shifting to the step 2 is carried out. Compared with the traditional convolutional neural network algorithms, the method of the invention carries out a weighted operation of spatial information multi-dimensions on a feature graph layer to compensate for a defect of spatial information losses caused by one-dimensional vectorization of the fully-connected layer.

Description

technical field [0001] The invention relates to an underwater target feature extraction method. Background technique [0002] At present, there are mainly two methods for underwater target feature extraction: time domain and frequency domain. The time domain is extracted from the waveform structure features. Reflected in the shape of the echo, the more obvious the target difference is, the more obvious the difference in the waveform structure is; in addition, the difference in the receiving angle of the echo and the attitude of the target will also have a greater impact on the waveform in the time domain, and these differences also hide the difference between the target. The characteristics of the object, that is, the classification features of the target are extracted from the waveform structure. Frequency domain features refer to the spectrum features obtained after signal processing. The target is identified by the method of spectrum estimation, and the target feature pa...

AI Technical Summary

Problems solved by technology

[0007] The feature information in two-dimensional images is often highly concealed, because changes in the position of the observation target and different observation angles will cause deformation, displacement or even distortion of the observation target.

In this process, spatial information will be lost and spatial features cannot be restored in the SoftMax layer, which will affect the classification accuracy and indirectly affect the quality of feature extraction when the network is continuously fed back and adjusted.

Images