Construction And Recognition Of Underwater Target Spectral Feature Map Based On Feature Fusion Network

With the strengthening of human development and utilization of the ocean,countries carry out extensive research on underwater target recognition technology.The signal collected by sonar contains target and environmental noise,therefore feature extraction becomes a key technology for underwater target recognition.However,recognizing acoustic signals in noisy environments has become a major challenge in underwater target recognition field.The noise energy masks the spectral energy of weak targets,which makes feature extraction difficult.This research studies the problem of low accuracy of underwater weak target recognition,designs an underwater target recognition system with visualization and automatic feature extraction,The Feature Fusion Network(FFN)is proposed based on improved VGG-16 for feature extraction to construct a spectral feature map of underwater targets that separates the line spectrum and noise,which achieves highly robust underwater weak target recognition.The specific research contents are as follows:Firstly,the concept of time-frequency analysis and popular transformation methods of underwater acoustic signals are discussed,the expression of line spectrum features by constructed spectrogram based on Low Frequency Analysis Recording(LOFAR),Mel filter and Gammatone filter is studied.The influence of the spectral features of underwater weak targets on the recognition performance is analyzed,and an underwater target recognition system framework with visualization and automatic feature extraction is designed.Secondly,in order to achieve the semantic segmentation of line spectrum and noise,line spectrum enhancement is transformed into the edge detection task of LOFAR spectrogram,the line spectrum detection performance of classical operators,artificially designed features and multi-level feature learning models are compared through the Fscore of the edge map.According to the multi-level feature representation of the LOFAR spectrum in the feature pyramid,a feature fusion weight assignment method of normalized F-score is proposed in the late fusion model,which improves the performance of line spectrum detection based on Holistic-Nested Edge Detection(HED),a spectral feature map with clean background was constructed in HED and part of weak line spectral enhancement was achieved.Then,aiming at the problem of missing detection of some weak line spectrum edges in late fusion model HED,FFN is proposed based on improved VGG-16 to construct the spectral feature map.FFN uses element-wise add and concatenate to add skip-layer connections to all convolutional layers of VGG-16,which obtains richer line spectrum features than HED.At the same time,FFN divides VGG-16 into five stages,the line spectrum edge localization accuracy is improved compared to HED by adding depth supervision to each stage.Further,a learning-based fusion weight allocation mode is designed for FFN,which improves the adaptability to the dataset compared to HED.The spectral feature map constructed by FFN achieves accurate semantic segmentation of background noise and target line spectrum.Finally,by cascading Soft-max classifiers after FFN,an underwater target recognition system with visualization and automatic feature extraction is built.In this system,the LOFAR spectrum with low signal-to-noise ratio is sent to FFN,and the spectral feature map produced by fusion output is sent to Soft-max as the result of feature extraction for classification.The classification results verify the effectiveness of feature extraction based on FFN.FFN enhances the line spectrum features and suppresses background noise,which greatly improves the recognition accuracy of underwater targets under the condition of low signal-to-noise ratio.