单步卷积神经网络在手机碎屏检测中的应用  

Application of One-step Convolutional Neural Network in Detection of Mobile Phone Broken Screen

作　　者：赵兴文 童飞 ZHAO Xingwen;TONG Fei(Department of Information Technology,Zhejiang Financial College,Hangzhou,China,310018)

机构地区：[1]浙江金融职业学院信息技术学院,杭州310018

出　　处：《福建电脑》2022年第12期25-29,共5页Journal of Fujian Computer

基　　金：浙江省金融教育基金会课题基金(No.2022Y36);浙江金融职业学院青年科研项目课题基金(No.2022YB44);浙江金融职业学院青年科研项目课题基金(No.2021YB33)资助。

摘　　要：针对工业生产生活中的手机碎屏检测需求,本文提出一种RandomMask的随机性数据增强策略。首先将较大像素的手机碎屏图像特征进行裁剪和缩放来获得备选mask,其次将训练图片进行四等分得到四个中心锚点,最后将备选mask随机覆盖任一锚点之上以实现增强图像特征表达的目标。同时本文采用基于卷积神经网络的单步目标检测架构的改进设计,聚焦解决工业质检、手机碎屏险自助理赔等服务在实时场景下的速度和精度要求,在网络输入端Input、中间网络Neck、骨干网络Backbone以及输出网络Head部分分别应用不同深度学习检测器组件得到本文最优算法模型。该模型配合RamdomMask图像增强策略可以实现以44fps的实时检测速度获得对于手机碎屏数据的87.9%mAp的平均精度。According to the requirements of mobile phone broken screen detection in industrial production and life,the paper proposes a random data enhancement strategy named RandomMask.Firstly,the broken screen features of larger pixels are cut and scaled to obtain the alternative mask.Secondly,the training image is quartered to obtain four central anchor points,and then the alternative mask is randomly covered on any anchor point to achieve the goal of enhancing the expression of image features.At the same time,the paper adopts the improved design of one-step object detection architecture based on convolutional neural network,focusing on solving the speed and accuracy requirements of industrial quality detection,mobile phone broken screen insurance self-service claim settlement and other services in real-time application.Different deep learning detector components are applied to the Input of the network,the Neck of network,the Backbone of network and the Head of network to obtain the optimal model of the paper,Combined with RamdomMask image enhancement strategy,the model can achieve an average accuracy of 87.9%mAp for mobile phone broken screen data at a real-time detection speed of 44fps.

关 键 词：卷积神经网络 深度学习 手机碎屏 碎屏检测 目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]