Childhood obesity is an escalating global health concern, emphasizing the need for early and effective detection methods. Traditional assessment techniques for fat distribution are often invasive and unsuitable for pediatric populations. This study explores a non-invasive alternative by integrating infrared thermal imaging (IRT) with deep learning models for classifying obesity levels among children. Thermal images were captured from the abdominal and neck regions of 150 children categorized as normal, overweight, or obese. A patch-based image analysis was performed using three models: a custom convolutional neural network (CNN), DenseNet121, and Inception ResNetV2. The custom CNN, developed and trained from scratch, consistently outperformed the other models, achieving accuracy rates of 93.1% for the abdomen and 91.6% for the neck, with balanced classification across all BMI categories. The abdominal region showed better discriminatory patterns, supported by higher AUC scores. An ablation study further validated the network’s robustness and highlighted the influence of hyperparameter optimization. The results indicate that deep learning applied to thermal imaging is a promising, non-invasive tool for early obesity screening in children.