Light-Deeplabv3+: a lightweight real-time semantic segmentation method for complex environment perception.

Current semantic segmentation methods have high accuracy. However, it has the disadvantage of high computational complexity and time consumption, which makes it difficult to meet the application requirements in complex environments. To achieve fast and accurate semantic segmentation of images, we propose a lightweight semantic segmentation method called Light-Deeplabv3+. First, a MobileNetV2Lite-SE architecture with SE module is proposed as the backbone network of the model, which can reduce the number of model parameters and improve the segmentation speed. Second, we propose an ACsc-ASPP module based on asymmetric dilated convolution block (ADCB) and scSE module to solve the semantic information loss during feature extraction. Our improvements can obtain more semantic features and improve segmentation accuracy. Finally, we propose a DSC-Blaze module to replace the original 3 × 3 standard convolution. It consists of depthwise separable convolution (DSC) and Blaze module, which can improve the model segmentation speed while maintaining the receptive field. The experimental results prove that the Mean Intersection over Union (MIoU) of Light-Deeplabv3+ on the PASCAL VOC2012 dataset is 73.15 % , and the parameter size is only 6.291MB. Its calculation amount is only 20.883G, and the speed on the 3060Ti platform is 37.66 frames per second (FPS). Compared with traditional Deeplabv3+, Light-Deeplabv3+ can achieve efficient and accurate image segmentation results with less computational overhead, and its performance is comparable to state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]