Efficient Lightweight Cnn Model For Real-time Semantic Segmentation On Autonomous Driving

Semantic segmentation is one of the most important challenges in computer vision which works to assigning each pixel of the input image to a semantic class.It also plays a significant role in practical systems as like autonomous driving and robot navigation vision.Recently,semantic segmentation has got to attention of many researchers which have shown numerous techniques using the convolutional neural networks(CNNs),where current approaches have achieved wonderful results for semantic segmentation but they have a large number of parameters and incur high computing costs.Designing an efficient real-time semantic segmentation model with competitive segmentation accuracy as well as a small size model is important.In this thesis,a novel efficient lightweight network architecture is proposed which called(ELNet)for real-time semantic segmentation and inspired by Res Net.It consists of two sections,the encoder network and the decoder network.The first section includes the main process encoder,which is a new residual block of Asymmetric-Dilated-Bottleneck(ADB)that utilizes asymmetric dilation convolutions with Point-wise and skip connections.These enable it to decreases the number of parameters and computational cost while maintaining the accuracy as much as possible and effective use.In the second section,the decoder,a Feature Attention Pyramid Network(FAPN)collects the feature maps of diverse scales to optimize performance through the feature attention pyramid mechanism.This confirms that the proposed model has a wide receptive field while keeping the model small-scale and shallow.Our proposed network achieves end-to-end and pixel-to-pixel training without pretraining from the scratch.In addition,our model achieves 73.3%,66% mean Io U and FPS 126 on Cityscapes,Camvid respectively with fewer than 0.97 M parameters,30 x fewer than those of others like Seg Net.We have evaluated ELNet on public self-driving datasets City Scapes and Camvid.Moreover,we have evaluated our results on NVIDIA TITAN XP GPU.