Multi-scale Model Integration And Optimization For Image Segmentation

In recent years,with the rapid development of the fields of automatic driving,image segmentation technology has become a research hotspot.In order to achieve better segmentation performance,a large number of pixel-level annotations are needed in practical applications.But it is time-consuming and labor-intensive to make such sample data.Considering that the public dataset of the generic target class already has some pixel-level annotation,the migration learning method can be utilized to improve the problem of lack of supervision information.Howerver,image segmentation based on such methods does not consider the diversity of target scales.And there exists the poor generalization performance of a single model and the bad image segmentation result caused by the wrong category prediction.Therefore,this paper uses the weakly supervised migration learning model as the basic framework of segmentation,constructs three multi-scale features networks to form differentiated homogenous base models,then applies the integrated approach to fuse the different base models,and finally improve the performance of the model using category prediction and credibility optimization.The main work is as follows:(1)An image segmentation method for learning multi-scale features is constructed.The overall segmentation framework still uses the codec structure of the weakly supervised migration learning model,and the part extracted by the encoder is replaced with a newly constructed pyramidal parallel input,nested hierarchical fusion,pyramidal pooled structure,and multi-scale features are extracted.The experimental results show that the pyramidal parallel input is more complete in processing the target detail information,and the hierarchical difference feature fusion is more smooth on the target edge contour segmentation,and the overall performance is comparable to that of the original migration learning algorithm.(2)An image segmentation method based on multi-scale model integration is proposed.Based on the multi-scale individual learner with certain accuracy and difference,the method uses the integrated algorithm to combine the parallel calculation results of each base model by weighted average meth od,and effectively avoids the difference by complementing each model.Performance shortcomings of a single model to enhance model generalization performance.The experimental results show that the performance of the dual model integration algorithm is 54.7%,the performance of the three-model integrated algorithm is 54.8%,and the original migration learning algorithm is increased by 5.2%.(3)An image segmentation algorithm based on fusion class prediction and credibility optimization is proposed.Based on the multi-scale integrated image segmentation model,this method separately evaluates the class prediction and pixel-level classification of image segmentation tasks,and introduces a classifier to give the target class information of the prediction process,so as to alleviate the segmentation caused by the class information error.The failure situation,combined with the credibility of the class and the pixel,effectively circumvents the false positive segmentation region.Finally,the average overlap rate of the algorithm on the VOC 2012 verification set is 58.3%,the average overlap rate on the test set is 57.5%.Compared with the original migration learning model,it is improved by 11.9%and 12.3%respectively.And it performs favorably against other segmentation methods using weakly-supervised information based on category labels as well.