The invention relates to an image feature learning method and device. The method comprises the steps of obtaining a feature vector of each training image through a neural network; determining a targetfunction value according to the matching degree of the feature vector of each training image and the feature vector recorded by each memory unit; updating the feature vector recorded by the corresponding memory unit according to the feature vector of each training image; and if the target function value does not meet the training target, optimizing the parameters of the neural network according to the difference between the target function value and the training target. According to the method, data annotation is not needed, it does not need to assume that input data obeys certain probabilitydistribution, and therefore flexibility is high, the calculation process is simple, implementation is convenient, and the learning effect can be guaranteed.