A Joint Dynamic Domain Adaptation Module and Multi-incentive-aware Based Model for Sentiment Analysis of Autism Paintings

With the development of society, the number of autistic children is increasing, and there is growing concern about their education and development. Autistic children often have difficulties in language expression, social interaction and emotional understanding. In this context, finding new ways to understand the emotions and needs of autistic children has become particularly important. By analyzing the drawings of autistic children and using deep learning models to mine the emotions and information embedded in them, we can provide educators and parents with a more intuitive and in-depth way to understand the inner world of autistic children. As a non-verbal form of expression, paintings can help autistic children release emotions and express their inner thoughts and feelings. Firstly, based on psychological research results, we propose a new image emotion classification network based on multi-incentive awareness, which consists of three phases, i.e., incentive selection (S), feature extraction (O) and emotion prediction (R). In the first phase, in this paper, we select special emotion excitations (color, object, face) by combining existing target detection methods in deep learning. Subsequently, we design three specific networks to extract different emotion features from multiple incentives synchronously, i.e., global network, semantic network, and expression network. Finally, the model combines all the above features for image emotion prediction. Comparative experiments demonstrate that the accuracy of our constructed sentiment analysis model outperforms existing algorithms. After training the above multi-incentive-aware image sentiment analysis model, we design a transfer learning model to map the features learned by the model on the regular image dataset to the drawing dataset of autistic patients. We perform multi-representation feature extraction using a hybrid structure, which consists of multiple sub-structural feature extraction units. At the same time we use several different sub-feature spaces to represent the original image emotional features and perform joint dynamic domain adaptation for domain data in each sub-feature space. Comparative experiments demonstrate that the accuracy of our constructed transfer learning model outperforms existing algorithms.