RCNN螺栓目标识别.docx

研究生周工作总结表（每周）时间实验件：白色平板；螺栓：黑%外六角5个，银色外六角5个，随机置于板上:本周具体进展情况测试：RCNN螺栓目标识别:网格选取：alexnet训练集构建：可寸，图间Lr；江；"：01p902>p119O3.pngOS4N>9O64m907.png08.pngO9.png10n9H4x*9标注:久图IlLJIl2:训练；在学GPU上训练。正在初始化融入敖婚归一化。轮I送代I经过的时词I小批更搅失I小批里窟硒度I小批里RMSEIRPNM.ni-baIChI1I1I00:00:03I2.2592I34.90%I0.15I19.20*I5IS0I00:01:16I1.4025I83.3I0.14I95.31等占GPUislowonae*ory.vhchcanslovperforaanceduetoadditionaldatatransfersthsain*eaory.fryreducntheIliniBatchSise,trainingIoption.Thisamngvi11notappearagainunlessyourunthecoaandearningon''MeKCnnrrnn£GPuLo>0neQT3)I10I100I00:02:36I1.15808488%I0.13I99.22%I15I150I00:03:54I1.0947I84.90«I0.13I100.00%I20200I00:05:10I0.9668I8733%I0.14I100.00¾测试:螺栓边缘霍夫线识别:输入：botl.pngbeh2p0boklp09WMpogbd602bopngbdt7nQbokll4M¼9boti2p9bohllpogbokl4ngboh1pc9boMlApog重整图片大小为256*256：中值滤波，去除孤立噪声点，保留边界信息。Canny边缘检测；由于边缘检测参数难以设定，编写小dem。GUI界面调整Canny参数，观察输出：（低阈值为0,高阈值150）£3cannydemoMinth.ld:0随动调整发现10左右较优BSnnydemoMinth.ld:10对底板螺栓不同色时检测效果更好:霍夫线检测：参数设定：最小线长度，最小检测点间距，最小角敏感度等输出：（进行角度计算最少需要条线）角度计算：点差计算初始角（保留负值）IQangle列表（List）（6元素）一I空值r0float1-47.3532968661031J1float1-47.3532968661831S2float115.3762512488261923float174.40718973366=4float115.945395900922856G5float175.96375653207353'u.Jj«mJrzV,Uj3、/jy-/r闲I关闭，JJe角度标准化：（不同线角差理应为120°）QIl室类型大小值C12float1float1float112.64670313389169112.64670313389169115.376251248826192C345float1float1float114.47189673365715.9453959O92285615.963756532073532r访,H3B最终输出角取均值:Anglefloat6414.497666501723272对应检测图: