心理健康是人们愈发关注的领域，随着我国经济的不断发展，国家也更加重视民众的心理健康问题；而灾难对人们心理的影响非常严重且持续时间较久，尤其是儿童和青少年群体，在较小的年纪受到巨大的冲击，比如过度惊吓、身体受到重伤、家人受伤或去世等，往往难以承受，常会出现一些心理问题。对于灾后心理健康问题的研究也在逐渐完善，本研究将优化灾后心理健康测评体系，使用树木绘画投射测验，并结合深度学习技术实现对画树测验进行绘画特征自动识别与图像分类，完成画树测验自动化评估模型的构建。

本次研究的施测地点为四川省雅安市芦山县，选择一所县初级中学，对全校约820名学生进行施测。研究一对学生作答的PTSD、抑郁、焦虑自评量表数据进行分析，结果表明其信效度良好；同时根据拟定的树木绘画测验特征评估指标对画树测验进行评估编码，再结合被试的量表得分筛选出量表分数差异显著的绘画特征；最后将选出的绘画特征作为特征变量，量表分数作为因变量，采用随机森林回归分析，进一步筛选出对被试群体心理问题的预测起到重要作用的绘画特征。研究二采用目标检测ssd_mobilenet_v1模型对研究一最终筛选出的绘画特征进行目标检测训练，对标注的绘画特征均能正常识别，平均精确率良好。研究三改进Inception v3图像分类模型，增加可输入绘画特征目标检测结果的分支，从而使模型在对画树测验分类时更加关注目标检测提取到的关键特征，进而提升分类准确度，增加模型的可解释性。

Mental health is an area of increasing concern, with the developing of our economy, the government also pays more attention to people's mental health problems. And the impact of the disaster on people's mental health is very serious and far-reaching, especially children and adolescents are greatly impacted at a young age, such as excessive shock, serious bodily injury, family member injury or death, they are often overwhelmed and suffer psychological problems. The research on post-disaster mental health problems is also improving. We will optimize the evaluation system of post-disaster mental health. This study will use the tree-drawing test, and combine with the deep learning technology to realize the automatic recognition of painting characteristics and classification of the tree-drawing test, and complete the construction of the automatic evaluation model of the tree-drawing test.

The study was conducted in Ya’an City, Sichuan Province, and selected a junior high school to test about 820 students. In research one, the data of PTSD, depression and anxiety self-assessment scale were analyzed, and the results showed that their reliability and construct validity were good, and tree-drawing tests was evaluated and coded according to the proposed characteristic evaluation indexes. Then, the painting features with significant differences in scale scores were selected. Finally, the selected painting features are taken as characteristic variables, the scale score is used as a dependent variable, and random forest regression analysis was used to further screen the important drawing characteristics for the prediction of psychological problems of the subjects. In research two, the ssd_mobilenet_v1 model of objection detection was used to conduct target detection training on the painting features finally screened in research one, and the marked painting features could be recognized normally with good average accuracy. Research three improves the Inception v3 image classification model, by adding the branches that can input the detection results of painting features, the model pays more attention to the key features extracted from target detection when classifying tree-drawing tests, thereby improving the classification accuracy and increasing the interpretability of the model.

The study found that there are some differences painting characteristics between the high-risk group of students in the disaster area and the normal student group. At the same time, it was also found that some subjects did not show any abnormalities in the responses in the scale, while there were more abnormal features in the pictures of trees. Therefore, it is feasible to use the drawing projection test as an auxiliary screening tool for mental health after the disaster. In addition, this study not only realized the object detection and classification evaluation of the tree-drawing tests, but also preliminarily tried to increase the interpretability of the classification process for the image classification model, laying a foundation for the improvement of the automatic evaluation of the tree-drawing test in the future.