Binary-SegNet: Efficient Convolutional Architecture for Semantic Segmentation Based on Monocular Camera

The paper utilizes a semantic segmentation model for images obtained from a monocular camera based on Binary Neural Networks (BNN). The proposed model uses Resnet18 as the backbone along with a decoder based on the PSP-Net architecture. By decreasing the magnitude of processing parameters and propagation within the network, this approach facilitates enhanced training efficiency, rapid inference speed, and reduced hardware demands. When implemented, our methodology produces superior outcomes in comparison to conventional CNN architectures. More precisely, the model’s inference pace experiences a 3.11-fold increase, whereas the accuracy remains untouched at 6.21%. Moreover, BNNs enable the implementation of semantic segmentation models on hardware that is resource-efficient or has a limited physical capacity. In conclusion, this architectural design presents prospects for integration into intelligent mechatronic systems, particularly in the development of knowledge systems for robotics traversing diverse environments, by maximizing costs and optimizing the monocular camera's operational capabilities.