The tourist landscape perception study of historic town is helpful for understanding the tourist's perception logic to the town's heritages and spaces from the perspective of 'subject-object interaction', so as to provide guidance for the heritage preservation and spatial renewal. Based on the principle of 'word bag' model and fully convolutional networks for semantic segmentation in computer graphics, this paper constructs a 'bag of photographic words' analysis method for public landscape perception study of historic towns. The method based on thematic cluster analysis of internet big image data, takes image dictionary construction, word bag generation, preference expression, perception law analysis as implementation steps, and extracts six landscape elements of layout, street, architecture, environment, customs and product, and also six image attribute characteristic describing elements which are original-derivative, natural-artificial, material-immaterial; overall-partial, concise-complex, vitality-quiet; photographing, recording, documentary were proposed for specific analysis, and use era, season, photographer's travel mode and age as the grouping labels for the bag of words research. Taking the Guizhou Qingyan Town as a sample, the paper analyzes various kind of situation and crowds, and finds that the public perception of the historic town landscape has the laws of 'perception degradation' and 'subject-object interaction', and the perception characteristics are affected by different experience pattern and age factors. Based on the above conclusions, the article further proposes flowing scenes setting measures for the historic towns that respond to the tourists' perception law from people-oriented perspective, so as to provide reference for relevant studies and practices.