<p>Indoor scene segmentation is an important research problem due to its widespread applications in the navigation of autonomous systems like robots, augmented reality experience, automation in smart home systems and healthcare assistive technologies for the elderly. In this paper, a pruned UNet architecture based on Particle Swarm Optimization (PSO) has been proposed and validated for indoor scene segmentation. Specifically, PSO was applied for each hidden layer of the network with an n-dimensional position vector that represents the number of kernels in the hidden layer. The accuracy on the test set serves as the fitness function. The model is progressively pruned by removing neurons corresponding to zero elements in the position vector. Applying PSO to all hidden layers in this fashion leads to an optimized UNet model with fewer neurons after each training epoch. The proposed UNet + PSO approach scores excellent results on the public benchmark ADE20K dataset, achieving training accuracy of 98% and testing accuracy of 63.83%, significantly outperforming standard UNet (32% training and 31.24% testing accuracy) and the FCN model (82.87% training accuracy and 55.04% testing accuracy). These results were calculated on a subset of ADE20K dataset from its bedroom scene images folder. Qualitative analysis of the predicted ground truths indicates the superiority of the UNet + PSO predictions over the FCN model.</p>