The invention discloses an improved method based on an extreme learning machine (ELM) and sparse representation classification. The method comprises the following steps of 1, randomly generating a hidden layer node parameter; 2, calculating a hidden layer node output matrix; 3, according to a size relation of L and N, using different formulas to calculate an output weight connecting a hidden layer node and an output neuron; 4, calculating an output vector of a query picture y; 5, determining a difference value of a maximum value of and a secondary maximum value os in an ELM output vector o, and if the difference value is greater than a set value, determining an index corresponding to the maximum value in the output vector, wherein the index is a type which the query picture belongs to; otherwise, entering into step6; step6, using a training sample corresponding to the k maximum values in the output vector o to construct a dictionary, using a coefficient reconstruction algorithm to calculate a linear representation coefficient of the picture y, calculating a residual error and determining the type which the query picture belongs to according to the type corresponding to the residual error. In the invention, a calculated amount is greatly reduced, a high recognition rate is realized and calculating complexity can be greatly reduced.