INKIND

BSc (Hons) Artificial Intelligence and Data Science | Data Science Group Project

Artificial Intelligence & Machine Learning
Child Psychology
Children's Drawings
Sanida Nethmitha VidanagamaSanida Nethmitha Vidanagama
Madiba Amindi lidiya Emily De Zoysa Gunathilaka RajapakseMadiba Amindi lidiya Emily De Zoysa Gunathilaka Rajapakse
Ediriyan Dhanuge Sanuli Yehansa DhanugeEdiriyan Dhanuge Sanuli Yehansa Dhanuge
Kaviyan RatneswaranKaviyan Ratneswaran

INKIND is a multimodal AI system designed to support the interpretable analysis of psychological and developmental indicators expressed in children’s drawings. Targeting children aged 2 to 11, the system provides educators and parents with structured, non-diagnostic insights, complementing professional evaluation rather than replacing it. The pipeline begins when a teacher uploads a child’s drawing alongside an optional description. A Vision Language Model first processes the raw drawing to generate two outputs: a child-perspective narrative and a comprehensive visual description of the drawing’s content. The teacher is then prompted to validate this AI-generated description before analysis proceeds, ensuring a human checkpoint at the earliest stage. The validated description and drawing are passed through an image normalisation pipeline, incorporating SAM-based segmentation, perspective correction, CLAHE contrast enhancement, and colour normalisation to standardise real-world photograph variations. The normalised image and the generated child-perspective description are then fed jointly into a multimodal mood classification model combining ResNet50 image features and DistilBERT text embeddings through early fusion, producing a Happy/Sad prediction with a confidence score. In parallel, a Drawing Indicator Analysis module which is powered by an LLM within a Retrieval-Augmented Generation framework grounded in psychological literature, uses the comprehensive visual description to extract and interpret structured drawing indicators such as line pressure, spatial usage, figure completeness, and shading intensity. A rule-based recommendation engine then maps the mood prediction, drawing indicators, and their interpretation into supportive, non-clinical activity suggestions delivered to the teacher through an Integrated Analysis Report.