| Mirrors are very common and important in daily life.In images containing mirror,objects are usually presented both inside and outside the mirror,which may interfere the ability of computer vision algorithms,such as object detection and image understanding.Mirror detection and segmentation is one of the ways to solve this problem.However,existing methods usually suffer from the problem of huge parameters,high difficulty of network training and high cost of resource limited device deployment.Due to the limited parameters and computation costs,the knowledge distillation based lightweight network has received wide attention in resourceconstrained algorithm modeling and it has exhibited its potential in image classification tasks.In this paper,we explore the ability of the kind of approach in optimizing the ability of mirror segmentation.However,directly applying the traditional knowledge distillation method for mirror segmentation may suffer from two problems:1)it is difficult to effectively realize knowledge transfer and correction;2)it is difficult to reasonably realize knowledge distillation if source data is unavailable.Therefore,this paper conducts the following researches.With respect to the first problem,we propose a mixed constraint knowledge distillation based mirror segmentation method.Firstly,a ‘teacher-student-discriminator’knowledge distillation model is developed by combining the teacher network,student network and adversarial discriminator network.Secondly,an improved lightweight student network is proposed based on Mobile Net V3-small network to enhance its perception ability on spatial contextual semantic information by adding a spatial information compensation module to the key layers.Finally,on top of the traditional knowledge distillation method,a mixed constrained objective function is constructed by introducing a rectification loss and an adversarial loss to improve the knowledge transfer ability and the correction ability towards the edge details.To verify the effectiveness of the proposed method,we carry out experimental validations on the MSD and PMD datasets.The results show that the proposed method has achieved some progresses on segmentation accuracy,computation and storage costs.With respect to the second problem,we propose a self-supervised knowledge distillation based mirror segmentation methods.On top of the ‘teacher-studentdiscriminator’ model,we explore how to conduct knowledge distillation by using unlabeled training data when the source data is unavailable and further propose a selfsupervised knowledge distillation method by using an unsupervised loss function and a self-adaptive loss function.The method can effectively utilize the unlabeled data in the domain to transfer knowledge from the teacher network to the student network,and can adaptively deal with the error propagation problem of the teacher network during the transfer process to improve the robustness of knowledge distillation.In order to validate the effectiveness of the proposed method,we carry out experimental studies on PMD and self-organized extended datasets.The results show that the proposed method can still improve the performance of mirror segmentation.Finally,a B/S architecture based prototype system is developed for mirror image segmentation by integrating the studied lightweight networks. |
