移动端实时数字识别检测

移动端实时数字检测

最近在做数字检测与识别，主要检测识别手机端的数字，设备为树莓派3。

我在树莓派上安装了Tesseract-OCR，发现检测数字的速度并不是很快，一串数字识别大概要3-400多毫秒，配上特定环境代码检测的话，至少要1s左右，这个速度要想达到实时检测是不可能的，因为要想用摄像头达到实时检测的效果，每帧识别至少要低于40ms，所以完全不够，最近用最简单的深度学习方法识别，在树莓派上，速度能达到2-3ms的速度！！不经感叹，CNN真的厉害。当然这还不够，因为这套算法识别，在手机上很受外部环境的干扰，比如关照，除非你在一个黑盒子，或者晚上来做检测。

 

这里先说一下识别达到的效果，比如手机上固定位置的一串数字，识别到了，树莓派上GPIO能做出响应，这样就能在手机上实现一些监测功能。

代码和效果图稍后更新。。。。。

楼主正在做另一种方法，避免关照的干扰，代码和效果图可能会更新晚一点。有兴趣做数字检测的同学可以一起交流。

 

部分代码

---

string modelTxt = "./lenet_train_test.prototxt";
		string modelBin = "./lenet_solver_iter_10000.caffemodel";
		//read image  

		dnn::Net net;
		try {
			net = dnn::readNetFromCaffe(modelTxt, modelBin);

		}
		catch (cv::Exception &ee) {
			cerr << "Exception: " << ee.what() << endl;
			if (net.empty()) {
				cout << "Can't load the network by using the flowing files:" << endl;
				cout << "modelTxt: " << modelTxt << endl;
				cout << "modelBin: " << modelBin << endl;
				return -1;
			}
		}
		//imshow("camera-demo", frame);
		blur(frame, frame, Size(3, 3));
		cvtColor(frame, gray_image, CV_BGR2GRAY);
		
		equalizeHist(gray_image, gray_image);
		//Rect first = selectROI("Test iamge", gray_image);
		//Mat ROI = gray(Rect(first.x, first.y, first.width, first.height));
		threshold(gray_image, result, 0, 255, THRESH_BINARY | THRESH_OTSU);
		imshow("gray_image", result);
		char c = waitKey(30);
		if (c == 27)
		{
			break;
		}
		vector > contours;
		vector hierarchy;
		findContours(result, contours, hierarchy, CV_RETR_EXTERNAL, CHAIN_APPROX_SIMPLE, Point());
		//cout << "contours" << contours.size() << endl;
		for (int i = 0; i < contours.size(); i++)
		{
			Rect rect = boundingRect(contours[i]);
			double area = contourArea(contours[i]);

			//cout << "area" << area << endl;
			if (area > 100)
			{
				drawContours(frame, contours, i, Scalar(0, 0, 255), 1, 8, hierarchy);
				rectangle(frame, rect, Scalar(255, 0, 0), 1, 8);
				Mat roi = result(rect);
				resize(roi, roi, Size(28, 28));
				copyMakeBorder(roi, dst, roi.rows*0.1, roi.rows*0.1, roi.cols*0.1, roi.cols*0.1, BORDER_CONSTANT, Scalar(0));
				TickMeter t;
				Mat inputBlob = dnn::blobFromImage(dst, 0.00390625f, Size(28, 28), Scalar(), false);
				Mat pred;
				net.setInput(inputBlob, "data");//set the network input, "data" is the name of the input layer   
				t.start();
				pred = net.forward("prob");//compute output, "prob" is the name of the output layer   
				t.stop();
				//cout << "pred" << pred << endl;
				int classId;
				double classProb;
				getMaxClass(pred, &classId, &classProb);
				cout << "Best Class: " << classId << endl;
				cout << "Probability: " << classProb * 100 << "%" << endl;
				ostringstream ss;
				ostringstream tt;
				ss.str("");
				tt.str("");
				ss << classId;
				tt << (int)t.getTimeMilli() /(int) t.getCounter();
				String prenum(ss.str());
				String time(tt.str());
				String label1 = "n:"+prenum ;
				String label2 = "time:" + time+"ms";
				putText(frame, label1, Point(rect.x, rect.y), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(255, 0, 0));
				putText(frame, label2, Point(rect.x, rect.y-15), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(255, 0, 0));
				//string k = to_string(i);
				//imshow(k, roi);
				imshow("final", frame);
			}
		}