In this study, we propose a method to clarify the application limits of DNN (Deep Neural Network) based visual inspection systems. A process for determining inspection results is a black box because DNN automatically extracts features. However, visual inspection requires judgment based on specifications. The problem is that the basis for the decision is unclear. To address this problem, we interpret the feature space of DNN using known features. Firstly, it generates data with explicit knowledge characteristics (for example, defect length, area, shading depth etc.) that can be arbitrarily modified. Secondary, the generated data are input to trained DNN models and observed the coordinate changes caused by their explicit knowledge characteristics in the feature space. Preliminary experiments on the MNIST dataset (public dataset) confirm that the DNN feature space is able to represent quantitative feature variation. Experiments using actual inspection images also confirmed the effectiveness of the proposed method.