A Study Of The Lightweight Weakly Supervised Semantic Segmentation Of Image

Semantic segmentation of images is one of the most important research branches in the field of computer vision,and also one of the key technologies of image processing and understanding.Semantic segmentation can be defined as the pixel classification task,which divides the pixels belonging to the same object in the image into the same category,and finally the image is divided into several disjoint areas to achieve the semantic level of object segmentation.Semantic segmentation has been widely used in various industries,such as medical image diagnosis,auto-drive system,remote sensing land monitoring and UAV image processing.However,current semantic segmentation methods cannot meet the practical application requirements.First of all,the training of neural networks depends on a large number of accurately labeled data,which undoubtedly increases the training cost.Secondly,the deep neural network has a large number of parameters,which demands huge memory space and computing ability.These issues make it difficult to directly deploy models on intelligent devices.In addition,human intervention in the design of network architecture leads to a low degree of automation,which does not conform to the concept of general artificial intelligence.Finally,the generalization performance of most segmentation models is poor,which requires manual parameter adjustment and full training for specific tasks,and it is difficult to quickly adapt to new tasks.Weakly supervised semantic segmentation has attracted much attention due to its low level of supervision and low difficulty of label acquisition.Weak labels used in the weakly supervised semantic segmentation can be divided into bounding box,scribble,point and image-level label.Among them,image-level label has the weakest level of supervision,which brings great challenges to the segmentation accuracy of the model.In view of the above problems,this work focuses on the weakly supervised semantic segmentation task based on image-level labels,and carries out research on the supervision level,lightweight and generalization performance of the deep models to achieve a better balance between the model efficiency,generalization and segmentation accuracy.Main innovative research results obtained in this work are as follows:(1)Given that the training of semantic segmentation network is limited by the number of accurate labels and memory resources,a lightweight weakly supervised semantic segmentation method based on clustering quantization is proposed.First,the double class activation mapping method is used to generate pseudo masks with higher accuracy to improve the issue of inaccurate object coverage caused by class activation mapping.Secondly,in the training process of the weakly supervised semantic segmentation network,the image parameters are lightened based on K-means clustering and reduce the memory occupation.Finally,the channel-spatial attention mechanism is introduced to improve the segmentation performance of the model by filtering important content and location information.The comparison and ablation experiments have demonstrated that this method achieves superior segmentation performance with lower training costs.(2)Aiming at the problems of the huge parameter quantity of neural network parameters,the high computational complexity and long training time,a lightweight method based on feature distillation is proposed to realize the tasks of weakly supervised semantic segmentation and semi-supervised image classification.For weakly supervised semantic segmentation,the pseudo masks generation method proposed by the previous chapter is improved at first.We propose a gradient-based double class activation mapping method,which avoids the restrictions of network structure on class activation mapping.Only several real labels are used for semisupervised classification task training.In the feature distillation module,the feature mapping from teacher network to student network is established to improve the feature extraction ability of small student network,thus improve the performance of lightweight model.In addition,it encapsulates a variety of classic backbone networks,and flexibly selects different network structures to participate in training by setting hyperparameters.Through image classification and segmentation experiments,it has been proven that this method can achieve good performance using lightweight networks.(3)For the weakly supervised semantic segmentation task,a lightweight neural architecture search method based on block structure is proposed aiming at the problem of long training time and high hardware requirements of the current neural architecture search methods.The gradient based double class activation mapping method is used to generate pseudo labels and participate in the training of segmentation architecture search.The reinforcement learning is used to search the network block structure and operations,the training process is accelerated by sharing operation units and knowledge distillation.2000 rounds of experiments have been conducted,and the results show that this method achieved lightweight neural architecture search process and could quickly search for network architectures with good segmentation performance.(4)Consider that the generalization performance of semantic segmentation network is poor and cannot adapt to new tasks quickly,a weakly supervised few-shot semantic segmentation method based on meta-learning is proposed.A weighted double class activation mapping method based on channel-spatial attention mechanism is proposed in the pseudo masks generation stage,which fully extends the coverage area of the target object and improves the accuracy of the pseudo labels.In the semantic segmentation stage,the meta-learning method based on optimization is adopted to quickly adapt to new tasks using prior knowledge and improve the generalization performance of the network.It avoids retraining the whole network every time and effectively improves the network efficiency.The experimental results show that the proposed method can effectively generalize to new tasks and obtain accurate semantic segmentation results.