Convolutional sparse representations express a signal by a sum of convolutional filters and corresponding sparse coefficients. This paper proposes to employ the multi-scale, namely variable size, filters to extract features for image classification. The experimental results on the Yale B Face Database show that the proposed method increases the accuracy by up to 4% compared to conventional method using single-scale convolutional sparse representations which is the filter size is fixed in the image classification with convolutional sparse coding classification (CSCC).