Research on Complex Scene Recognition Based on Semantic Segmentation

Image semantic segmentation plays an important role in scene understanding, which is a hot topic in the field of computer vision. There are two main methods to improve the semantic segmentation accuracy of complex scenes in the existing models. One is to consider the spatial relationship between pixels, but it will produce a high amount of calculation. The other is to expand the receptive field, but the context representation is still not clear enough. Aiming at the semantic segmentation problem in complex scenes, a convolution neural network segmentation model with attention mechanism is proposed to detect the recognition technology of traffic scenes. Context Semantic Encoding(CSE) module is introduced to capture the global context information and highlight the category information associated with the scene. Multi-scale feature extraction is realized to increase the weight of the foreground target feature. The distribution law between autonomous learning data of generative confrontation network can solve the problem of ignoring the spatial relationship between pixels. The network is implemented based on PyTorch framework. The experimental results on the Cityscapes dataset show that the mloU reaches 75.7 %.