Aiming to the poor retrieval accuracy of many existing image retrieval methods,in this paper,color moments,multi-block local binary patterns,gray leve l co-occurrence matrix, scale-invariant feature transform and spatial location information of an image are calculated to extract five types of feature that represent the essential characteristics of an image from d ifferent perspectives,then,five sparse representation classifiers are constructed based on these five types of feature respectively according to the labeled information of each training image in the database.Meanwhile,the decision fusion strategy is introduced,and the fusion weights of these five classifiers are adaptively determined by an iterat ive procedure according to the different classification performance of each sub-classifier.For the fusion wei ghts can judge the capability of different types of feature for describing the image′s property,the distance cor rection factors are constructed based on the fusion weights to correct the distances between the images described by d ifferent types feature respectively. Finally,the image similarity on the basis of these features is calculated to ob tain the retrieval results.The proposed method is evaluated on the Corel-1000image dataset,and the experimental resul ts demonstrate that the average precision rate can reach 82.1% which increases by about 10% compared with that o f the state-of-the-art methods,and the proposed method can recall images with higher accuracy and robustness than compared methods.