Abstract: ‍ ‍This paper addresses the issue of inadequate utilization of local channel information in traditional object recognition networks by proposing a featured channel grouping attention mechanism. This mechanism is integrated with residual convolutional neural networks to create an effective feature extraction network. Initially， the features are segmented along the channel dimension， resulting in multiple sub-features. Within these sub-features， the significance of each channel is highlighted， and appropriate weights are assigned. Channel rearrangement is then applied to form sub-feature groups， facilitating enhanced information exchange among the overall channels. Following this， the average pooled feature map of the sub-features is utilized as a representative， allowing for further information exchange to enhance and amalgamate both the overall and local channel information of the features. To further enhance the recognition performance of the network， this paper uses the Low-Frequency Analysis and Recording （LOFAR） spectrum and the Mel spectrum of underwater acoustic target radiation noise as inputs for the network model. It constructs a feature fusion network using an autoencoder to achieve information exchange between different features. The time-frequency features of the two input signals are deeply fused to improve the feature representation of the information conveyed by the signal. Experimental validation using the ShipsEar dataset shows that the improved attention mechanism proposed in this paper increases recognition accuracy by over 1.38% compared to commonly used channel attention mechanisms. The fusion of the two features for recognition enhances accuracy by 6.17% and 1.2%， respectively， compared to using the LOFAR and Mel spectra separately.