The invention discloses a non-reference screen image quality evaluation method based on non-supervision feature learning; the method comprises the following steps: obtaining a normalization screen image of a plurality of distortion-free screen images; using ZCA operations and a non-supervision clustering algorithm to obtain a dictionary code book according to the normalization screen image; obtaining a normalization screen image of an evaluated distortion screen image; using a Gaussian kernel similarity weight method and a K-nearest Neighbor method to process the normalization screen image, thus obtaining a weight feature matrix; using an LLC algorithm to obtain an LLC feature vector according to the dictionary code book and the weight feature matrix; finally using the support vector regression technology to test the LLC feature vector, and predicting to obtain an objective quality evaluation prediction value of the evaluated distortion screen image. The method can fully consider the influences of local information changes on the visual quality, thus improving the correlation between the objective evaluation result and the subjective perception.