With increasing research on image feature vector extraction and classification, blind steganalysis is becoming more efficient and accurate. However, many existing methods use similar processing for all images without taking account the diverse image contents. This paper proposes a new approach based on
image contents and feature fusion. The input images are divided into several classes according to the content complexity before feature extraction. Bhattacharyya distance is used to evaluate the usefulness of individual features and determine their weights. Steganalysis is subsequently conducted using a fusing approach and a support vector machine (SVM) classifier in a decision making process. Experimental results on several sets of  images demonstrate that the proposed steganalyzer outperforms some previous methods. It provides reliable results with reduced computational complexity.