Real Time Semantic Segmentation Algorithm based on Encoder-decoder

Abstract

Encoder-decoder networks have demonstrated significant advantages in semantic segmentation tasks. PP-LiteSeg as a light-weight encoder-decoder network, has gained a lot of attention because it can strike a good balance between accuracy and speed. However, the encoder and decoder in the networks have obvious limitations in extracting semantic information and recovering detailed information, respectively. To address these limitations, we carry out some improvements on PP-LiteSeg and propose MFF-LiteSeg. The improvements include a new encoder based on a parallel dilated convolution short-term dense cascade module and a multi-scale context module. The former module are exploited to help the encoder capturing rich features while maintaining low parameter counts. The latter module is integrated into the encoder to capture multi-scale contextual information, enhancing the ability of the encoder to understand objects of different scales. The improvements also include an attention fusion module which is applied to the decoder to optimize the feature fusion, enabling the decoder to more accurately restore detailed feature information. Experimental results on public datasets demonstrate that MFF-LiteSeg achieves higher segmentation accuracy than PP-LiteSeg while maintaining a fast inference speed of 160 FPS, striking a good balance between inference speed and segmentation accuracy.