Research And Application Of Image Semantic Segmentation Algorithm Based On CNN

As a key technology of computer vision,image semantic segmentation has important research value.In order to solve the shortcomings of traditional methods,make the target more accurately extracted from the background,and combine deep learning with image semantic segmentation to have a better understanding and analysis of the target.Image semantic segmentation is mainly to classify the pixel level of the image and then segment it into regions with different semantics.Due to the rapid development of convolutional neural networks,neural networks capable of autonomously analyzing and learning image features can be obtained by using images and annotations to train neural networks,thereby greatly improving the accuracy of semantic segmentation.Deep learning has wide application potential in the field of image semantic segmentation research,so it is of great practical significance to carry out research on semantic segmentation algorithms based on convolutional neural networks.In view of the poor performance of the single model in the image segmentation technology,the problems of poor image recognition accuracy and category prediction failure will cause the semantic segmentation effect to be relatively rough.In this paper,the global context structure module and multi-scale feature fusion module are used and the improved semantic segmentation method is adopted.Then,a CNN-based relative quality prediction network is used to test the consistency of objective indicators and subjective judgments through a regression model,so as to evaluate the performance of the semantic segmentation algorithm and verify the advancedness and reliability of the algorithm through experiments.By applying the improved model to public datasets such as PASCAL and Cityscapes,experimental research shows that the introduction of global context structure into image semantic segmentation can improve the MIo U value by 0.71% and the PA value by 0.36%.Furthermore,after combining the selected decoder module and the improved multi-scale module,the above-mentioned key values have been increased by 2.05% and 0.38%,respectively.The global context structure can improve the accuracy of the test without increasing a large amount of test time.The multi-scale feature fusion algorithm can improve the accuracy of the algorithm,but it will affect the time performance.Compared with the benchmark model VGG16 and the residual network,the accuracy of this model has been improved to a certain extent,which basically meets the real-time requirements.Through the quality prediction network model,it can be seen that the results are consistent with the subjectively evaluated performance indicators and have good robustness.