Image Semantic Segmentation Algorithm Based on HN-DeepLabV3 +

Image semantic segmentation algorithm divides the image into several specific regions with unique properties, and extracts the objects of interest, it has been widely used in medical image analysis, intelligent transportation system, automatic driving and other fields. Aiming at the problems of incomplete boundary segmentation, background interference, easy missed detection, false detection and incomplete feature extraction in today 's image segmentation algorithms, a HN-DeepLabV3+ semantic segmentation algorithm is proposed. In the encoder, the MobileNetV3 lightweight module is used instead of Xception as the feature extraction backbone network to improve the segmentation efficiency of the image. The pooling pyramid combining the dilated spatial convolution of multiple receptive fields and the mixed dilated convolution is used to reduce the feature loss in the feature extraction process. The attention module NAM is introduced to improve the accuracy of the segmentation results. The experimental results show that the HN-DeepLabV3+ algorithm has high accuracy and robustness in image segmentation tasks. The average pixel accuracy of the improved algorithm is 84.15 %, the average intersection ratio is 72.18 %, and the average pixel accuracy of the analogy is 82.54 %. Compared with the original DeepLabV3+ algorithm, it is increased by 3.93 %, 2.88 % and 0.42 %, respectively.