Due to the inherent characteristics of sonar imaging principles and the interference of complex underwater environments， underwater sonar images generally suffer from insufficient resolution and missing target details. To address these issues， a high-quality sonar image generation method based on multi-scale feature fusion was proposed. Firstly， the Residual Dense Blocks （RDBs） were used to extract image features at shallow level， thereby capturing basic texture and contour information， and establishing spatial layout of the image. Secondly， a Multi-Scale Attention feature extraction module （MSA） was designed to focus on key features at different scales adaptively and further enhance the expression of key features while suppressing redundant information expression through the attention mechanism. Finally， a discriminator network was constructed using a pixel-by-pixel discrimination strategy based on spectral normalization， which improved the reconstruction ability of complex object contours and details. Experimental results on an underwater sonar image dataset show that the proposed method achieves relative improvements of 6.7% and 5.4%， respectively， in Peak Signal-to-Noise Ratio （PSNR） and Structural SIMilarity （SSIM） metrics compared to the existing representative method ESRGAN （Enhanced Super-Resolution Generative Adversarial Network）. It can be seen that the proposed method improves the generation performance on underwater sonar image dataset effectively.

Focusing on the issue that most unsupervised feature selection models based on dictionary learning cannot fully exploit the intrinsic correlations among data， which reduces the accuracy of feature importance judgment， an unsupervised feature selection model with Dictionary Learning and Sample Correlation Preservation （DLSCP） was proposed. Firstly， the original data were encoded by learning the dictionary atoms， and the latent representations to characterize data distribution were obtained in the dictionary space. Secondly， the intrinsic correlations among data were learned adaptively in the dictionary space to alleviate the influence of redundant and noisy features， thus obtaining accurate local structure among data. Finally， the intrinsic correlations among data were used to measure the relevance and importance of data features. Experimental results on TOX dataset show that， when selecting 50 features， DLSCP improves the Normalized Mutual Information （NMI） and clustering Accuracy （Acc） by 13.33 and 7.95 percentage points respectively compared to non negative spectral analysis model NDFS（Nonnegative Discriminative Feature Selection） and by 15.74 and 7.31 percentage points respectively compared to unsupervised feature selection model with hidden space embedding LSEUFS （Latent Space Embedding for Unsupervised Feature Selection via joint dictionary learning）， which verifies the effectiveness of DLSCP.

In the Case-Based Reasoning (CBR) case retrieving and matching, different cases are usually composed by different features. But most of the traditional CBR engines adopt fixed feature weights mode, which makes matching rate of whole system very low. To solve this problem, this paper proposed a CBR engine model with variable feature weights and brought interactive mode into feature weights calculating module. It calculated subjective weight based on group decisionmaking theory and proposed an adjustment method which used differences between a single expert and his group. It used similarity rough set theory to calculate objective weight in order to make results calculating more objective and accurate. At last, it designed composite weights adjustment algorithm which calculated the distance between the subjective weight and objective weight, considered the deviation degree of those two weights, then deduced weights adjustment coefficient, and get the final weight adjustment results. The calculation example and simulation analysis of network attack cases validate the effectiveness of the proposed method and prove this method has much better performance in different performance indexes.