In the past, CNN trains the model with back-propagation. The model is lack of explanation and has large quantity of computation, so a CNN without back-propagation (FF-CNN) is proposed recently. The model replaces the convolution part with feature extraction method based on PCA. But PCA inputs the training data in a vector form. For images, it loses the information between different order so that the performance is limited. This study proposed a classification model called Pixel-Anchored CNN (PA-CNN) which modifies the FF-CNN and replaces PCA stage with the High-Order Principal Component Analysis (HOPCA). It reduces quantity of computation and the loading of memory and the performance slightly increases.