Scholarlink

<p>Automatically generating descriptions for House-Tree-Person (HTP) drawings can effectively reduce the subjective components in clinical assessments and assist clinical psychologists in better understanding the psychological state of the evaluated individuals. This process significantly shortens the substantial time and human resources required for manual evaluation of HTP drawing tests. Current image captioning algorithms have demonstrated remarkable effectiveness in extracting features from images and generating corresponding textual descriptions. However, models trained on public datasets are not well-suited for psychological analysis of HTP drawings. In this paper, we have constructed a Chinese dataset for HTP psychological analysis research (HTPICC1k), which collects and annotates a significant number of images related to HTP psychological analysis. Additionally, to address the limitation that current image captioning algorithms can only generate brief image descriptions, we propose an Image-Title-Caption Composer (ITCC) to enhance the quality of image descriptions. First, features of house, tree, and person in HTP drawings are extracted. These features are used to generate corresponding titles, and specific image descriptions are then generated based on these titles. Experimental results demonstrate that ITCC trained on HTPICC1k can automatically generate psychological analysis descriptions related to HTP drawings, and it outperforms ClipCap, BLIP, and BLIP2 models in multiple evaluation metrics, with CIDEr and BLEU-4 scores of 91.02 and 41.04, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Psychological analysis of house-tree-person drawings based on multimodal large models

  • 期刊论文

  • 发表日期：

    2025年1月8日

查看全文
  • Dahong Xu,
  • Siyu Jiang,
  • Yihan Zhang,
  • Xi Li

摘要

Automatically generating descriptions for House-Tree-Person (HTP) drawings can effectively reduce the subjective components in clinical assessments and assist clinical psychologists in better understanding the psychological state of the evaluated individuals. This process significantly shortens the substantial time and human resources required for manual evaluation of HTP drawing tests. Current image captioning algorithms have demonstrated remarkable effectiveness in extracting features from images and generating corresponding textual descriptions. However, models trained on public datasets are not well-suited for psychological analysis of HTP drawings. In this paper, we have constructed a Chinese dataset for HTP psychological analysis research (HTPICC1k), which collects and annotates a significant number of images related to HTP psychological analysis. Additionally, to address the limitation that current image captioning algorithms can only generate brief image descriptions, we propose an Image-Title-Caption Composer (ITCC) to enhance the quality of image descriptions. First, features of house, tree, and person in HTP drawings are extracted. These features are used to generate corresponding titles, and specific image descriptions are then generated based on these titles. Experimental results demonstrate that ITCC trained on HTPICC1k can automatically generate psychological analysis descriptions related to HTP drawings, and it outperforms ClipCap, BLIP, and BLIP2 models in multiple evaluation metrics, with CIDEr and BLEU-4 scores of 91.02 and 41.04, respectively.