【作者】 王海波；

【导师】 叶宾；

【作者基本信息】 中国矿业大学 ， 控制工程（专业学位）， 2021， 硕士

【摘要】 近些年来,深度学习在计算机视觉领域得到了广泛的应用。特别是随着计算机硬件算力的飞速发展,越来越多的行业都进行了智能化产业升级,极大优化了产品的生产流程和应用场景。目前,深度神经网络已广泛应用于智能医疗、自动驾驶、智慧交通、智能安检等场景。实例分割作为一个综合性图像处理任务,在这些行业中起到了巨大的作用。实例分割统一了分类、检测、分割三大视觉任务,是对图像中的实例进行像素级别的识别,是一种更符合人类视觉认知的建模方法。目前已有的实例分割方法结构复杂,计算量大,不能达到实时处理,然而对于大部分应用场景,模型的推理速度至关重要。因此本文针对目前存在的问题,通过对模型的训练过程及网络结构进行优化,设计出能够实时运行的高精度单阶段实例分割方法。本文遵循单阶段方法的思路,以SOLO v2模型框架为基础,设计更高效的实例分割方法。本文主要完成了以下工作:(1)为了给模型训练提供更多高质量的正样本,本文设计了适用于实例分割任务的自适应样本分配策略ASA。通过实验证明,ASA策略可以在不增加模型推理时间的前提下有效提高实例分割方法的分割精度。(2)本文基于SOLO v2模型框架提出了高性能单阶段实例分割方法E-Seg,通过设计增强特征金字塔模块来丰富特征的多样性,以及递进式掩码生成网络来提高原型掩码的质量。E-Seg在保证推理速度的前提下,大大提高了实例分割的精度,最终E-Seg在COCO 2017验证集上的精度为36.6mAP,在1080TI硬件上的推理时间为32ms。(3)本文基于Pytorch框架设计了实例分割工具箱FlySeg,将实例分割模型的搭建过程模块化,通过搭积木的方式构建模型,大大提升了算法研究效率。FlySeg中复现的多种模型的精度与官方公布的精度差距均低于0.5mAP,证明了FlySeg的有效性。更多还原

【Abstract】 In recent years,deep learning has been widely used in the field of computer vision.Especially with the rapid development of computer hardware computing power,more and more industries have carried out intelligent industrial upgrading,which greatly optimizing the production process and application scenarios of products.At present,deep neural network has been widely used in intelligent medical,automatic driving,intelligent transportation,intelligent security inspection and other scenes.As a comprehensive image processing task,instance segmentation plays a huge role in these industries.Instance segmentation unifies the three visual tasks of classification,detection and segmentation,it can recognize the instances in the image at the pixel level,which is more in line with human visual cognition.At present,the existing instance segmentation methods have complex structure and large amount of computation,it can not achieve real-time processing.However,the inference speed of the model is very important for most application scenarios.Therefore,in view of the existing problems,this thesis designs a real-time high-performance single-stage instance segmentation method by optimizing the training process and network structure of the model.Following the idea of single-stage method,this thesis designs a more efficient instance segmentation method based on the SOLO v2 framework.This thesis mainly completes the following works:(1)In order to provide more high-quality positive samples for model training stage,this thesis designs an adaptive sample assignment strategy ASA for instance segmentation task.Experiments show that the ASA strategy can effectively improve the segmentation accuracy without increasing the inference time.(2)This thesis proposes a high-performance single-stage instance segmentation method E-Seg based on SOLO v2 framework.This thesis designs an enhanced feature pyramid module to enrich the diversity of features,and a progressive mask generation network to improve the quality of the prototype mask.E-seg improves the accuracy of instance segmentation while ensuring the inference speed.Finally,the accuracy of ESeg on COCO 2017 validation set is 36.6 m AP,and the inference time on 1080 TI is 32 ms.(3)This thesis proposes an instance segmentation toolbox FlySeg based on Pytorch.FlySeg modularizes the process of building the model.We can design the model by building blocks,which greatly improves the efficiency of algorithm research.The accuracy difference between the reproduced models in FlySeg and the officially published models is less than 0.5 m AP,which proves the effectiveness of FlySeg.更多还原