Multi-resolution Neural Network for Road Scene Segmentation

This study presents a deep neural network for semantic segmentation in road scene images. We have used the convolution neural network in the encoder-decoder format. The architecture consists of convolution layers, max-pooling layers, batch normalization layers, and ReLU as nonlinear layers in contracting and expanding paths. This study uses the Cambridge-driving Labeled Video Database (CamVid) dataset for implementation. The multi-resolution convolution blocks in the proposed neural network enable hierarchical feature learning and contextual understanding. In scenarios like road scenes, where the objects are of different sizes, the multi-resolution convolution blocks provide the adaptability to label the pixels efficiently. We trained the networks with controlled hyperparameters. Also, we compare the performance of our proposed neural network with the existing state-of-the-art models. Hausdorff distance is noted as 199.69 ± 31.83, 204.82 ± 28.78, 202.21 ± 30.92, 203.57 ± 32.39, 195.46 ± 30.12, 200.37 ± 31.48 and 135.36 ± 23.34 for Pretrained SegNet, SegNet, VGG16, MobileNetV2, Pretrained DeepLabv3+, Pretrained FCN and the proposed model, respectively. Hausdorff’s distance of the proposed model emphasizes efficient delineation of object boundaries, and the overall quality of the segmented images is analyzed using mean Intersection of Union (mIoU), mean accuracy, and mean BF score. Observation shows that the proposed network is robust in segmenting the less frequent classes and class imbalance objects.