Research And Application Of Image Semantic Segmentation Algorithm Based On Deep Neural Network

Image semantic segmentation refers to the category prediction of each pixel in the image,dividing the different categories by using different colors,while also retaining the position information of the pixels of each category in the picture,which is one of the core tasks of image understanding.The Deeplab v3+ model is a fully supervised coding-decoding model structure with better performance.Its complex coding structure can effectively extract the high-level semantic features of the image,but only one-scale coding features are merged during the sampling process of the feature map.This decoding method will cause the loss of multi-scale contextual details in the encoding process,making the final segmentation result rough.Fully supervised semantic segmentation models require a large number of manually annotated images,while existing semi-supervised models have problems with high computational complexity and large memory consumption.Aiming at the above problems,this paper mainly improves from two aspects,and applies the improved model to the remote sensing image dataset Vaihingen to further verify its generalization performance.(1)An image semantic segmentation algorithm based on multi-scale feature adaptive fusion(ASFF-Net)is proposed,which can effectively fuse multi-scale detail information in the coding process of Deeplab v3+ model,and use it for decoding and upsampling the encoded feature map.Experimental results on the common dataset Cityscapes and remote sensing image dataset Vaihingen show that the algorithm has certain effectiveness and is more accurate in segmenting small-scale targets.(2)A semi-supervised image semantic segmentation algorithm based on holistically-guided decoding is proposed,which uses the generative adversarial network to promote the prediction results to obey the label distribution;in the generator,the multi-scale contextual information in the encoding is captured by using the holistically-guided decoder to generate semantically rich high-resolution feature maps;at the same time,the discriminator is used to generate a confidence map for unlabeled data,determine the trusted area in the prediction results,and thus provide a supervisory signal for semi-supervised learning.Experimental results on the common dataset PASCAL VOC2012 and the remote sensing image dataset Vaihingen show that the model can achieve a semantic segmentation effect comparable to that of the current semi-supervised model with better performance,using only 1/3 of its computation amount and 1/2 memory usage.