<p>This article provides an in-depth analysis of the problem of automatic detecting and classification of human emotions based on facial images. The study examines the relevance of identifying emotional states through facial expressions, as well as the advantages and limitations of existing approaches. An approach based on the geometric features of the face is proposed for emotion recognition, in which a normalised feature space is constructed using 68 specific facial points. Based on these features, a multilayer perceptron (MLP) neural network was developed and trained using an optimised architecture. The study utilised widely used databases such as FER2013 and KDEF. At the face detection stage, the Haar Cascade algorithm was employed, while a regression tree–based model from the Dlib library was used to detect facial landmarks. The neural network model was trained and evaluated in the Python programming environment using the TensorFlow and Keras libraries. The experiments yielded an accuracy of 93% on the FER2013 database and 91% on the KDEF dataset for the proposed MLP model. The obtained results were compared with classical classifiers such as SVM, KNN, Random Forest and Decision Tree, confirming the superiority of the proposed approach. The research findings demonstrate that geometric facial features and lightweight neural network models are effective for emotion recognition based on facial expressions and open up possibilities for their application in real-time intelligent systems. Keywords: facial image, emotion, neural network, Haar classifier, feature point, Multi-layer perc eptron.</p>