Weakly Supervised Image Segmentation Based On Priori Informations And Its Applications

With the application of artificial intelligence in industrial and medical fields,how to segment the “intereste” areas in images has become a hot research topic.Most of the current image segmentation methods are strongly supervised,which are effective but need to be supported by labeled datasets.Since the weakly supervised image segmentation methods are not only has lower training cost than strongly supervised methods,but also has better effect than unsupervised methods,which can balance the algorithm effect and cost.Therefore,thesis mainly focuses on the a priori-based weakly supervised image segmentation algorithm,and the main research content is the following three parts:In the thesis,we propose a sparsity region segmentation algorithm based on generative model(SSBG)for the problem that artificially generated target regions can lead to overfitting of weakly supervised algorithms.A sparsity loss function is designed for sparse target regions as a regular term to reduce the overfitting of the model.The generative sub-network learns the distribution of the background images and completes the image reconstruction.The discriminant sub-network measures the distance between the target region and the background,which can also effectively reduce the overfitting problem.In test stage,SSBG compares the reconstructed images with the input images,and the regions with larger differences are identified as target regions.The experimental results of15 categories on MVTec AD show that the method outperforms other superior generative models in most categories in terms of AUROC and AP both outperform other superior generative modeling schemes,with an average AUROC of 93% and an AP of 71.8%.In the thesis,we propose a shape-based image segmentation network(SBSN)to address the problem that complex shape priors are difficult to express explicitly.SBSN uses a convolutional restricted Boltzmann machine to explicitly express the complex shape prior by learning.First,the Boltzmann machine is pre-trained to obtain a basic shape prior.In the training stage,the shape prior is added to the loss function to guide the segmentation network training so that the output shape distribution is close to the pre-trained shape prior.Meanwhile,the Boltzmann machine continues to learn the shape prior using the current samples.Experiments are conducted for four different targets: cracks,horses,airplanes,and people,and the average AUROC and ABO reach 92.8% and 93.1%,respectively.The effectiveness of the SBSN scheme and the shape prior is verified by comparative experiments.In the thesis,we combine sparsity and shape prior to develop a screen text segmentation system and apply the system to medical screens.The screen text not only occupies a relatively small portion of the image,but also has a relatively fixed shape,which satisfies the applicability conditions of both sparsity and shape prior.According to business requirements,the system is divided into five modules: input module,pre-processing module,segmentation module,storage module and display module.The segmentation module uses the SBSN algorithm architecture and guides the network training using both the sparsity prior and the shape prior.A prototype of a text segmentation system for medical screens is built to reduce the workload of doctors by implementing it in a medical scenario.The feasibility and effectiveness of the system is confirmed by testing with real data.