add dll version and its usage
This commit is contained in:
UnknownObject
@@ -0,0 +1,263 @@
|
||||
#include "pch.h"
|
||||
#include "OCRConfig.h"
|
||||
#include "EasyOCR_Detector.h"
|
||||
|
||||
cv::Mat uns::EasyOCR_Detector::NormalizeMeanVariance(const cv::Mat& in)
|
||||
{
|
||||
cv::Mat img;
|
||||
in.convertTo(img, CV_32FC3);
|
||||
cv::Scalar mean(0.485f * 255, 0.456f * 255, 0.406f * 255);
|
||||
cv::Scalar var(0.229f * 255, 0.224f * 255, 0.225f * 255);
|
||||
img -= mean;
|
||||
img /= var;
|
||||
return img;
|
||||
}
|
||||
|
||||
void uns::EasyOCR_Detector::AdjustResultCoordinates(EOCRD_Rects& polys, float ratioW, float ratioH, float ratioNet)
|
||||
{
|
||||
for (auto& poly : polys)
|
||||
{
|
||||
for (auto& pt : poly)
|
||||
{
|
||||
pt.x *= ratioW * ratioNet;
|
||||
pt.y *= ratioH * ratioNet;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void uns::EasyOCR_Detector::ResizeAspectRatio(const cv::Mat& src, cv::Mat& dst, float squareSize, float magRatio, float& ratio, cv::Size& heatmapSize)
|
||||
{
|
||||
int h = src.rows, w = src.cols, c = src.channels();
|
||||
float target = magRatio * std::max(h, w);
|
||||
if (target > squareSize)
|
||||
target = squareSize;
|
||||
ratio = target / std::max(h, w);
|
||||
|
||||
int targetH = int(h * ratio), targetW = int(w * ratio);
|
||||
cv::resize(src, dst, cv::Size(targetW, targetH));
|
||||
|
||||
int h32 = (targetH + 31) / 32 * 32;
|
||||
int w32 = (targetW + 31) / 32 * 32;
|
||||
cv::Mat canvas = cv::Mat::zeros(h32, w32, src.type());
|
||||
dst.copyTo(canvas(cv::Rect(0, 0, targetW, targetH)));
|
||||
dst = canvas;
|
||||
heatmapSize = cv::Size(w32 / 2, h32 / 2);
|
||||
}
|
||||
|
||||
bool uns::EasyOCR_Detector::GetDetBoxesCore(const cv::Mat& textmap, const cv::Mat& linkmap, float textThresh, float linkThresh, float lowText, EOCRD_Rects& boxes, cv::Mat& labels, std::vector<int>& mapper, bool estimateNumChars)
|
||||
{
|
||||
cv::Mat tmap = textmap.clone(), lmap = linkmap.clone();
|
||||
int H = tmap.rows, W = tmap.cols;
|
||||
// 1. 二值化 & 合并
|
||||
cv::Mat textScore, linkScore;
|
||||
cv::threshold(tmap, textScore, lowText, 1, cv::THRESH_BINARY);
|
||||
cv::threshold(lmap, linkScore, linkThresh, 1, cv::THRESH_BINARY);
|
||||
cv::Mat combined;
|
||||
cv::add(textScore, linkScore, combined);
|
||||
combined = cv::min(combined, 1);
|
||||
// 2. 连通域
|
||||
int nLabels = 0;
|
||||
cv::Mat stats, centroids;
|
||||
cv::Mat combined8u;
|
||||
combined.convertTo(combined8u, CV_8U);
|
||||
try
|
||||
{
|
||||
nLabels = cv::connectedComponentsWithStats(combined8u, labels, stats, centroids, 4);
|
||||
}
|
||||
catch (cv::Exception e)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
// 3. 遍历每个 label
|
||||
for (int k = 1; k < nLabels; ++k)
|
||||
{
|
||||
int area = stats.at<int>(k, cv::CC_STAT_AREA);
|
||||
if (area < 10)
|
||||
continue;
|
||||
// 文本阈值过滤
|
||||
cv::Mat mask = (labels == k);
|
||||
double maxVal;
|
||||
cv::minMaxLoc(tmap, nullptr, &maxVal, nullptr, nullptr, mask);
|
||||
if (maxVal < textThresh)
|
||||
continue;
|
||||
// 构建 segmap
|
||||
cv::Mat segmap = cv::Mat::zeros(H, W, CV_8UC1);
|
||||
segmap.setTo(255, labels == k);
|
||||
mapper.push_back(k);
|
||||
// 删除 link 区域
|
||||
segmap.setTo(0, (linkScore == 1) & (textScore == 0));
|
||||
// 膨胀
|
||||
int x = stats.at<int>(k, cv::CC_STAT_LEFT);
|
||||
int y = stats.at<int>(k, cv::CC_STAT_TOP);
|
||||
int wbox = stats.at<int>(k, cv::CC_STAT_WIDTH);
|
||||
int hbox = stats.at<int>(k, cv::CC_STAT_HEIGHT);
|
||||
int niter = int(std::sqrt(area * std::min(wbox, hbox) / float(wbox * hbox)) * 2);
|
||||
int sx = std::max(0, x - niter), sy = std::max(0, y - niter);
|
||||
int ex = std::min(W, x + wbox + niter + 1), ey = std::min(H, y + hbox + niter + 1);
|
||||
cv::Mat kernel = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(1 + niter, 1 + niter));
|
||||
cv::dilate(segmap(cv::Rect(sx, sy, ex - sx, ey - sy)), segmap(cv::Rect(sx, sy, ex - sx, ey - sy)), kernel);
|
||||
// 轮廓拟合最小外接矩形
|
||||
std::vector<cv::Point> pts;
|
||||
cv::findNonZero(segmap, pts);
|
||||
cv::RotatedRect rect = cv::minAreaRect(pts);
|
||||
cv::Point2f boxPts[4];
|
||||
rect.points(boxPts);
|
||||
std::vector<cv::Point2f> box(boxPts, boxPts + 4);
|
||||
// diamond->rect
|
||||
float wlen = (float)cv::norm(box[0] - box[1]);
|
||||
float hlen = (float)cv::norm(box[1] - box[2]);
|
||||
float ratio = std::max(wlen, hlen) / (std::min(wlen, hlen) + 1e-5f);
|
||||
if (std::abs(1 - ratio) <= 0.1f)
|
||||
{
|
||||
int minx = W, maxx = 0, miny = H, maxy = 0;
|
||||
for (auto& p : pts)
|
||||
{
|
||||
minx = std::min(minx, p.x); maxx = std::max(maxx, p.x);
|
||||
miny = std::min(miny, p.y); maxy = std::max(maxy, p.y);
|
||||
}
|
||||
box =
|
||||
{
|
||||
{ float(minx),float(miny) },
|
||||
{ float(maxx),float(miny) },
|
||||
{ float(maxx),float(maxy) },
|
||||
{ float(minx),float(maxy) }
|
||||
};
|
||||
}
|
||||
// 顺时针起点
|
||||
int start = 0;
|
||||
float minSum = box[0].x + box[0].y;
|
||||
for (int i = 1; i < 4; i++)
|
||||
{
|
||||
float s = box[i].x + box[i].y;
|
||||
if (s < minSum)
|
||||
{
|
||||
minSum = s;
|
||||
start = i;
|
||||
}
|
||||
}
|
||||
std::rotate(box.begin(), box.begin() + start, box.end());
|
||||
boxes.push_back(box);
|
||||
}
|
||||
return (!boxes.empty());
|
||||
}
|
||||
|
||||
uns::EasyOCR_Detector::EasyOCR_Detector()
|
||||
{
|
||||
ort_inited = false;
|
||||
ort_cpu_session = nullptr;
|
||||
model_path = G_OCRConfig.GetDetectModelPath();
|
||||
ort = OrtGetApiBase()->GetApi(ORT_API_VERSION);
|
||||
}
|
||||
|
||||
bool uns::EasyOCR_Detector::Init()
|
||||
{
|
||||
if (ort_inited)
|
||||
return true;
|
||||
if (!RecheckModelInfo())
|
||||
return false;
|
||||
try
|
||||
{
|
||||
ort_env = Ort::Env(ORT_LOGGING_LEVEL_WARNING, "EasyOCR_Detector");
|
||||
bool fallback_to_cpu = false;
|
||||
if (!OCRToolBox::AutoSelectEP(ort, ort_session_options, fallback_to_cpu))
|
||||
return false;
|
||||
OCRToolBox::InitOrtSessionOptions(ort_session_options);
|
||||
if ((G_OCRConfig.GetGPUUsage() == easyocr::GPUUsage::CPUOnly) || fallback_to_cpu) //使用CPU则初始化cpu session
|
||||
{
|
||||
ort_cpu_session = new Ort::Session(ort_env, model_path.c_str(), ort_session_options);
|
||||
//通过CPU session获取输入输出名
|
||||
OCRToolBox::GetInputOutputNames(ort_cpu_session, input_names, input_ns, output_names, output_ns);
|
||||
}
|
||||
else
|
||||
{
|
||||
//通过临时session获取输入输出名（CUDA：线程不安全）
|
||||
Ort::Session ort_session(ort_env, model_path.c_str(), ort_session_options);
|
||||
OCRToolBox::GetInputOutputNames(&ort_session, input_names, input_ns, output_names, output_ns);
|
||||
}
|
||||
ort_inited = true;
|
||||
return true;
|
||||
}
|
||||
catch (...)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
bool uns::EasyOCR_Detector::UnInit()
|
||||
{
|
||||
try
|
||||
{
|
||||
if (ort_cpu_session != nullptr)
|
||||
delete ort_cpu_session;
|
||||
ort_cpu_session = nullptr;
|
||||
return true;
|
||||
}
|
||||
catch (...)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
bool uns::EasyOCR_Detector::RecheckModelInfo()
|
||||
{
|
||||
if (model_path.empty())
|
||||
model_path = G_OCRConfig.GetDetectModelPath();
|
||||
return OCRToolBox::CheckFile(model_path);
|
||||
}
|
||||
|
||||
uns::EOCRD_Rects uns::EasyOCR_Detector::operator()(const cv::Mat& image)
|
||||
{
|
||||
// 0. check model
|
||||
if (!RecheckModelInfo())
|
||||
return {};
|
||||
try
|
||||
{
|
||||
// 1. resize + normalize
|
||||
cv::Mat resized;
|
||||
float ratio;
|
||||
cv::Size heatmapSize;
|
||||
ResizeAspectRatio(image, resized, 1280.0f, 1.5f, ratio, heatmapSize);
|
||||
cv::Mat input = NormalizeMeanVariance(resized);
|
||||
// 2. 构造成 NCHW tensor
|
||||
cv::dnn::blobFromImage(input, input);
|
||||
std::array<int64_t, 4> shape = { 1, 3, input.size[2], input.size[3] };
|
||||
Ort::MemoryInfo memInfo = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
|
||||
Ort::Value tensor = Ort::Value::CreateTensor<float>(memInfo, input.ptr<float>(), input.total(), shape.data(), shape.size());
|
||||
// 3. ONNX 推理
|
||||
auto outputs = ((ort_cpu_session != nullptr) ? ort_cpu_session->Run(Ort::RunOptions{nullptr}, input_names.data(), & tensor, 1, output_names.data(), 1) : Ort::Session(ort_env, model_path.c_str(), ort_session_options).Run(Ort::RunOptions{nullptr}, input_names.data(), & tensor, 1, output_names.data(), 1));
|
||||
std::vector<int64_t> outputShape = outputs[0].GetTensorTypeAndShapeInfo().GetShape();
|
||||
// 4. 拆分 score_text & score_link
|
||||
float* outData = outputs[0].GetTensorMutableData<float>();
|
||||
int H = int(outputShape[1]), W = int(outputShape[2]);
|
||||
cv::Mat score_text(H, W, CV_32F);
|
||||
cv::Mat score_link(H, W, CV_32F);
|
||||
for (int y = 0; y < H; ++y)
|
||||
{
|
||||
for (int x = 0; x < W; ++x)
|
||||
{
|
||||
int offset = (y * W + x) * 2;
|
||||
score_text.at<float>(y, x) = outData[offset + 0];
|
||||
score_link.at<float>(y, x) = outData[offset + 1];
|
||||
}
|
||||
}
|
||||
// --- 3. 得到 boxes/polys (heatmap 坐标系) ---
|
||||
EOCRD_Rects boxes, polys;
|
||||
cv::Mat labels;
|
||||
std::vector<int> mapper;
|
||||
if (!GetDetBoxesCore(score_text, score_link, textThreshold, linkThreshold, lowText, boxes, labels, mapper, false))
|
||||
return {};
|
||||
polys = boxes;
|
||||
// --- 4. 计算最终映射比例 ---
|
||||
float invRatio = 1.0f / ratio;
|
||||
float ratioNetW = float(resized.cols) / float(heatmapSize.width);
|
||||
float ratioNetH = float(resized.rows) / float(heatmapSize.height);
|
||||
// 通常 heatmapSize = (resized.cols/2, resized.rows/2)，所以 ratioNetW/H ≈ 2
|
||||
// --- 5. 映射回原图 ---
|
||||
AdjustResultCoordinates(polys, invRatio, invRatio, ratioNetW);
|
||||
return polys;
|
||||
}
|
||||
catch (...)
|
||||
{
|
||||
return {};
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user