检测
使用R-CNN深学习检测器检测对象
描述
bboxes=检测(探测器,我)我使用一个R-CNN检测器与卷积神经网络(地区)对象。返回对象的位置检测的边界框。
使用这个函数时,使用CUDA®使英伟达®GPU是强烈推荐。GPU大大减少了计算时间。使用GPU的并行计算需要工具箱™。关于支持计算能力的信息,看到万博1manbetxGPU的万博1manbetx支持版本(并行计算工具箱)。
(___,还返回一个分类标签分配的数组边界框,使用前面的语法。标签用于定义对象类在训练使用标签)=检测(探测器,我)trainRCNNObjectDetector函数。
(___)=检测(___,检测对象在指定的矩形的搜索区域roi)roi。
(___)=检测(___,使用一个或多个指定选项名称,值)名称,值对参数。例如,检测(探测器,我,‘NumStongestRegions’, 1000)限制最强的地区建议1000的数量。
例子
火车R-CNN停车标志探测器
负荷训练数据和网络层。
负载(“rcnnStopSigns.mat”,“stopSigns”,“层”)
映像目录添加到MATLAB路径。
imDir = fullfile (matlabroot,“工具箱”,“愿景”,“visiondata”,…“stopSignImages”);目录(imDir);
设置网络训练选项使用mini-batch 32减少GPU内存使用量的大小。降低InitialLearningRate减少的速率网络参数改变。这是有益的在微调pre-trained网络和阻止网络改变过快。
选择= trainingOptions (“个”,…“MiniBatchSize”32岁的…“InitialLearnRate”1 e-6…“MaxEpochs”10);
火车R-CNN探测器。训练可以花几分钟来完成。
rcnn = trainRCNNObjectDetector (stopSigns层,选择,“NegativeOverlapRange”0.3 [0]);
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *培训R-CNN对象探测器以下对象类:* stopSign步骤1的3:从27日训练图像提取区域建议……。步骤2的3:在训练数据训练神经网络分类对象……| = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = | | | |时代迭代时间| Mini-batch | Mini-batch |基地学习| | | |(秒)| | |精度损失速率| | = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = | | 3 | 50 | 9.27 | 0.2895 | 96.88% | 0.000001 | | 100 | | 14.77 | 0.2443 | 93.75% | 0.000001 | | 150 | | 20.29 | 0.0013 | 100.00% | 0.000001 | | 200 | | 25.94 | 0.1524 | 96.88% | 0.000001 | | = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = |网络训练完成。步骤3的3:训练边界框回归模型为每个对象类……100.00%……做的。R-CNN培训完成。* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
测试R-CNN探测器测试图像。
img = imread (“stopSignTest.jpg”);[bbox,分数,标签]=检测(rcnn img,“MiniBatchSize”、32);
显示强大的检测结果。
[得分,idx] = max(分数);bbox = bbox (idx:);注释= sprintf (' % s:(信心= % f) '标签(idx),得分);detectedImg = insertObjectAnnotation (img,“矩形”bbox,注释);图imshow (detectedImg)
删除图像目录的路径。
rmpath (imDir);
