Accurate emotion recognition in real-world environments remains a persistent challenge due to diverse and unpredictable visual conditions. This paper explores the enhancement of emotion recognition under challenging real-world conditions, where conventional models frequently fail to achieve reliable performance. It investigates the impact of factors such as occlusions, lighting variations, poses, expression intensities, and image quality on the performance of deep learning models, utilising the BAUM-1 dataset to simulate these real-world scenarios. Modifications and filters were applied to the dataset, including various occlusions like sunglasses and blurred rectangles, as well as changes in illumination and image quality. A Convolutional Neural Network (CNN) was specifically adapted to address these real-world challenges. The model underwent training and testing across a spectrum of conditions, revealing variable accuracy levels in response to the different challenges, particularly noting a significant impact from occlusions. Despite this, the model showed a notable resilience against certain variations in illumination and occlusions.