SCNN

Use of convolutional neural networks for semantic image segmentation across different computing systems

“…However, these functions result in dense activations which are computationally expensive (Kernell 2018). Modern neural networks commonly use the Rectified Linear Unit (ReLU) function because it results in sparse and efficient activations making it less computationally expensive (Parashar et al 2017).…”

“…The VGG-16 model architecture is comprised of 16 weighted layers and is freely available for download online through GitHub. The VGG-16 is distinguished from other CNNs by the use of many convolutional layers with progressively smaller receptive fields, as opposed to only a few convolutional layers with large fields (Parashar et al 2017). The receptive fields in a CNN model represent the center area of focus (spatial extent) the model is looking at, as each convolution is applied.…”

“…This architecture was chosen because the model is simpler to train due to the reduced number of parameters and an increased number of nonlinearities in between the neuron layers allowing the model to efficiently learn and achieve better performance. This is important because it makes the decision function more discriminative between the different classes (Parashar et al 2017).…”

“…The performance of semantic segmentation systems can be evaluated based on metrics such as execution time, memory footprint, and accuracy (Parashar et al 2017). Since the majority of systems are constrained by the time that can be spent on the inference pass, speed or runtime is a particularly valuable metric.…”

“…In fact, even GPU-accelerated neural networks are often constrained by the number of GPU cores and memory capacity (Long et al 2014). Pixel accuracy is arguably the most important metric and generally used to gauge the performance of semantic segmentation techniques (Parashar et al 2017). In this work all three metrics are used to assess algorithm performance and segmentation accuracy.…”