Abstract: Focusing on the problem of natural image classification,we propose a novel semi-supervised multi-instance learning(MIL) algorithm based on fuzzy latent semantic analysis(LSA) and transductive support vector machine(TSVM).This algorithm regards each image as a bag,and the low-level visual feature of the segmented regions as instances.In order to convert MIL problem into a standard supervised learning problem,firstly,all the instances in training bags are clustered by K-Means method,and each cluster center serves as a "visual word" to build "visual vocabulary table".Secondly,according to the distance between "visual word" and instance,a fuzzy membership function is defined to establish a fuzzy "term-document" matrix,then LSA method is used to obtain bag's(image's) latent semantic models,which can converts every bag to a single sample.Finally,in order to use the unlabeled images to improve classification accuracy,the semi-supervised TSVM is used to train classifiers.Experimental results on the Corel image set show that compared with the traditional LSA method,the fuzzy LSA classification accuracy is increased by 5.6%,and its performance is superior to other MIL algorithms.