Image Segmentation And Saliency Detection Based On Activate And Multitask Learning

In recent years,the performance of deep learning models for segmentation is greatly driven by largescale training data with high quality pixel-wise segmentation labels.However,the annotation for segmentation is very labor intensive that the annotation has to be completed pixel by pixel.In this time-consuming and labor-intensive situation,we consider that the contribution of different images to the segmentation models may vary greatly,some images have a large gain on the model,but some images do not contribute to the gain of the model,that is,not all images should be annotated.It is challenging to make good use of the large-scale unlabeled images for the performance of semantic segmentation and annotate the images with the least labor cost.To this end,this paper proposes a segmentation framework for human-machine collaboration for binary classification.We use some binary detection datasets for saliency detection to prove our conclusions,so this work is closely related to saliency detection tasks.Therefore,there are some problems of inaccurate edge segmentation and sample imbalance in segmentation and saliency tasks.How to consider both inaccurate edge segmentation and sample imbalance problems,this paper proposes a strategy.The main contributions of this paper are as follows:(1)Deep learning technology is used in combination with active learning and progressive learning to form a complete framework for the annotation of segmenting data and models.Active learning strategies are used to determine the difficulty of samples to mine low-confidence samples.Progressive learning strategies are used to adjust hyperparameters to adapt to changes in the model during iterations.(2)Aiming at the problems of sample imbalance and insufficient segmentation edge accuracy,this paper proposes a multi-task loss joint optimization saliency detection algorithm.By combining the four loss functions BCE loss,Io U loss,SSIM loss,and F-measure loss,the performance of the model is improved significantly.