One of the most important situations in which data is used in corporate activities is in marketing activities, where it is important not only to analyze customer purchasing behavior but also to optimize the company's product lineup to meet the needs of consumers. Although many product position maps have been proposed in the past, all of them require human sensitivity data for each product, which is expensive to analyze, and there is no useful model that can evaluate multiple products efficiently and from a bird's eye view. Therefore, the objective of this research is to develop a low-cost technology to construct a deep learning model that evaluates the impression or image (emotional quality) that consumers have of product images, and to empirically demonstrate that this model can be a powerful tool for visualizing the emotional quality of multi-products and optimizing product lineups.In FY2023, we proposed a deep learning model for evaluating product images as the basis of our research, applied it to real data, and analyzed the results in a paper. Unlike conventional artificial intelligence approaches to images, which recognize objective concepts such as "dog" or "cat," this method is an artificial intelligence that evaluates images subjectively, such as "cute" or "vivid". While Kansei Engineering is a well-known analysis method for subjective sensitivity evaluation, it requires detailed analysis of all products and evaluation of combinations of products. Therefore, we proposed a method to evaluate the sensibility of all product images at a low cost using deep learning, and examined the method's effectiveness using actual product image data. We examined the effectiveness of the proposed method using actual product image data. The results showed that even when the number of evaluations is small, it is possible to create product position maps. In addition to the above research paper, we also studied improvement ideas to make the proposed method a more powerful tool, and presented the results both domestically and internationally. For example, regarding data selection for efficient model training, we showed that accuracy can be improved by obtaining ratings for as many different product images as possible, considering the fact that the data used are the results of pairwise comparisons between product images. We have also studied the results of combining multiple sensibility evaluation indices and confirmed that the accuracy of estimating the evaluation values of product images can be improved depending on the combination of indices.