Research On Image Classification Of The House-Tree-Person Drawing Based On CNN

Objective:Objective:we use the deep learning mature technology,convolutional neural network,solve the problem of subjectivity in evaluation of HTP drawing projection test in this research field,abandon the traditional painting experience of painting evaluation relies on the feature extraction and classification,the use of convolution HTP drawing image expression characteristics of automatic learning neural network model,make reflect on mental behavior of normal and abnormal classification of image painting.Methods:(1)we sampled 1074 primary school students in grade 5--8,completed the Achenbach self rating scale for adolescents,and completed the housing tree drawing test,in Hunan,Liuyang,Leiyang,Changsha and other places.According to the criterion of dividing the score between the developers of the scale and the related research scale in China,we divide the housing tree sample into two categories:normal and abnormal,and generate the supervised data set of house tree.(2)Use convolutional neural network model as the characteristics and classification of the image,the pixel matrix painting images as the input features(variables,feature),the classification results for the label(variable,label),the data set into the model,convolutional neural network after convolution,pool,fully connected,classification output such operations,from the bottom of the pixel matrix features to high-level semantic feature mapping,automatic learning to HTP drawing image features,the classification of normal and abnormal behavior of adolescent psychological HTP drawing images.Results:(1)based on the self rating scale of the demarcation standard behavior of Achenbach adolescents,divided into 1074 subjects,211 of them classified as abnormal samples,863 were assigned to the normal samples,the total abnormal rate was 10.4%,8 of at least one factor score is above a critical value of the students accounted for 20.58%of the total number of tests.(2)After several rounds of iterative convolution neural network model,the loss of value and the correct rate tends to be stable,the classification of the training set and test set the correct rate of around 0.90,the area under the ROC curve was 0.833,the classification of HTP drawing image has certain accuracy,the model has a better classification performance.Conclusion:(1)The rate of abnormal psychological behavior in the assessment of primary and middle school students in three areas of Hunan was in accordance with the other studies in Hunan.(2)without subjectively influenced by researchers,convolutional neural network models effectively classify the mental and behavioral characteristics of image representation of house tree painting,and apply convolutional neural network to the classification diagnosis of Fang Shu painting test.