Unsupervised　dimension　reduction　has　gained　widespread　attention.　Most　of　previous　work　performed　poorly　on　image　classification　due　to　taking　no　account　of　neighborhood　relations　and　spatial　localities.　In　this　paper,　we　propose　the　＇regularized　convolutional　auto-encoder＇,　which　is　a　variant　of　auto-encoder　that　uses　the　convolutional　operation　to　extract　low-dimensional　representations.　Each　auto-encoder　is　trained　with　cluster　regularization　terms.　The　contributions　of　this　work　are　presented　as　follows:　First,　we　perform　different　sized　filter　convolution　in　parallel　and　abstract　a　low-dimensional　representation　from　images　cross　scales　simultaneously.　Second,　we　introduce　a　cluster　regularized　rule　on　auto-encoders　to　reduce　the　classification　error.　Extensive　experiments　conducted　on　six　publicly　available　datasets　demonstrate　that　the　proposed　method　significantly　reduces　the　classification　error　after　dimension　reduction.