Road Scene Image Segmentation based on Deep Learning and Meta Pruning

Semantic segmentation of road scenes is crucial in autonomous driving technology, as it involves identifying and understanding the surrounding environment in real time and accurately, which requires a segmentation network that can process a large amount of image data quickly in real time and accurately distinguish different objects in a complex and changing environment, usually this model is too complex to guarantee real time or too simplified, resulting in a large reduction in accuracy. Existing lightweight models primarily focus on reducing computational load and the number of parameters, but the significance of network parameters to model performance is ignored or difficult to distinguish. To overcome the issue, we propose a novel approach combining deep learning with meta pruning to construct a light weighted road scene segmentation model, The advantage of meta pruning is that it evaluates the significance of different parameters to the network model, retaining the significant parameters and removing the minor ones to achieve precise pruning. The current results show that the method performs well on standard data sets.