blumia/deepin-ocr
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: 修复长条图片无法检测到文本 (#5)

Browse Source 
原因是当图片高度或宽度低于一定值后，检测算法难以执行检测
经过测试，图片的短边大于等于64的时候即可满足检测算法的要求
修改方法为在预先缩减图片尺寸后，进一步判断图片的短边是否符合大于等于64的要求，对不符合的执行第二次尺寸计算

Log: 修复长条图片无法检测到文本
This commit is contained in:
wzyforgit 2022-05-27 16:22:58 +08:00 committed by GitHub
parent b834d19c00
commit b7d77703f5
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 17 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
src/paddleocr-ncnn/details.cpp
View File

@ -34,7 +34,7 @@ std::vector<std::vector<std::vector<int>>> Details::detectText(const cv::Mat &sr
int w = src.cols;
int h = src.rows;
//输入图片固定尺寸960*960
//1.缩减尺寸
float ratio = 1.f;
if (std::max(w, h) > 960) {
if (h > w) {
@ -47,9 +47,23 @@ std::vector<std::vector<std::vector<int>>> Details::detectText(const cv::Mat &sr
int resizeH = int(h * ratio);
int resizeW = int(w * ratio);
resizeH = std::max(int(round(float(resizeH) / 32) * 32), 32);
resizeW = std::max(int(round(float(resizeW) / 32) * 32), 32);
//2.扩大尺寸，确保图片的最短的边大于等于64,小于这个值可能会导致检测不到文字
float ratio2 = 1.f;
if (std::min(resizeW, resizeH) < 64) {
if (resizeH > resizeW) {
ratio2 = 64.0f / resizeW;
} else {
ratio2 = 64.0f / resizeH;
}
}
resizeH = int(resizeH * ratio2);
resizeW = int(resizeW * ratio2);
//3.标准化图片尺寸，否则会错位
resizeH = int(round(float(resizeH) / 32) * 32);
resizeW = int(round(float(resizeW) / 32) * 32);
//执行resize，记录变换比例
cv::Mat resize_img;
cv::resize(src, resize_img, cv::Size(resizeW, resizeH));
float ratio_h = float(resizeH) / float(h);