Research On Image Semantic Segmentation Based On Weakly Supervised Learning

In recent years,semantic segmentation has made great progress in the fields of autonomous driving,medical image analysis,and geological detection.With the increasing computing power of computers,the real-time performance of the segmentation model can also be better guaranteed.These fields have achieved rich results based on fully supervised convolutional neural networks,but fully supervised neural networks need to use a lot of pixel-level labeling information,and pixel-level labeling information requires a lot of labor costs,so the semantic segmentation method based on weakly supervised learning has become a new research hotspot.The weakly supervised semantic segmentation method uses non-pixel-level weak tags such as image class labeling to achieve pixel-level segmentation of the image.Therefore,how to use weak information tags to find prominent target regions in the image is the key to the weakly supervised semantic segmentation method.The segmentation method based on the growth of the seed region is a way to solve the problem of weakly supervised semantic segmentation.This method first generates the seed region based on the class label,and then expands the seed region to form the final segmentation map.However,this method has the problems of noise category output and insufficient semantic information.Based on the above problems,this thesis proposes improvements to the existing methods.The main work of this thesis is as follows:(1)For the problem that the category activation mapping method can only generate sparse seed regions,this thesis proposes a multi-scale feature extraction model based on the multi-scale context acquisition method,and proposes a fusion method of heat maps at different scales.The heat map generated by the model can more effectively capture the target area and generate higher quality seed areas for subsequent models.(2)Based on the problem of insufficient noise category output and semantic information in the deep seed region-growing model,this thesis proposes a noise suppression branch and self-attention mechanism module to alleviate the above problems.The improved model can generate dense semantic segmentation masks with higher quality.The experimental results show that the improved scheme proposed in this thesis can effectively improve the segmentation performance of the model.(3)Using the segmentation mask generated by the previous model as pseudo-annotation information,the encoder-decoder segmentation network based on the fusion of multi-scale features is used to obtain the end-to-end segmentation network by retraining,and the model is validated on the data set.