Zinnia is a simple, customizable and portable online hand recognition system based on Support Vector Machines. Zinnia simply receives user pen strokes as a sequence of coordinate data and outputs n-best characters sorted by SVM confidence. To keep portability, Zinnia doesn't have any rendering functionality. In addition to recognition, Zinnia provides training module that allows us to create any hand-written recognition systems with low-cost.
% tar zxfv zinnia-X.X.tar.gz % cd zinnia-X.X % ./configure % make % su # make install
Install recognition models
% tar zxfv zinnia-tomoe-XXXX.tar.gz % zinnia-tomoe-XXXX % ./configure % make % su # make install
DLL and executable are included in zip file. No recognition models are included. Run the following command after putting model files in zip archived directory.
bin\zinnia_convert.exe handwriting-ja.model.txt handwriting-ja.model bin\zinnia_convert.exe handwriting-zh_CN.model.txt handwriting-zh_CN.model
Zinnia provides C/C++ libraries. Perl/Ruby/Python libraries are also available with SWIG
#include <stdio.h>
#include "zinnia.h"
int main(int argc, char **argv) {
size_t i;
zinnia_recognizer_t *recognizer;
zinnia_character_t *character;
zinnia_result_t *result;
recognizer = zinnia_recognizer_new();
if (!zinnia_recognizer_open(recognizer, "/usr/local/lib/zinnia/model/tomoe/handwriting-ja.model")) {
fprintf(stderr, "ERROR: %s\n", zinnia_recognizer_strerror(recognizer));
return -1;
}
zinnia_character_t *character = zinnia_character_new();
zinnia_character_clear(character);
zinnia_character_set_width(character, 300);
zinnia_character_set_height(character, 300);
zinnia_character_add(character, 0, 51, 29);
zinnia_character_add(character, 0, 117, 41);
zinnia_character_add(character, 1, 99, 65);
zinnia_character_add(character, 1, 219, 77);
zinnia_character_add(character, 2, 27, 131);
zinnia_character_add(character, 2, 261, 131);
zinnia_character_add(character, 3, 129, 17);
zinnia_character_add(character, 3, 57, 203);
zinnia_character_add(character, 4, 111, 71);
zinnia_character_add(character, 4, 219, 173);
zinnia_character_add(character, 5, 81, 161);
zinnia_character_add(character, 5, 93, 281);
zinnia_character_add(character, 6, 99, 167);
zinnia_character_add(character, 6, 207, 167);
zinnia_character_add(character, 6, 189, 245);
zinnia_character_add(character, 7, 99, 227);
zinnia_character_add(character, 7, 189, 227);
zinnia_character_add(character, 8, 111, 257);
zinnia_character_add(character, 8, 189, 245);
result = zinnia_recognizer_classify(recognizer, character, 10);
if (result == NULL) {
fprintf(stderr, "%s\n", zinnia_recognizer_strerror(recognizer));
return -1;
}
for (i = 0; i < zinnia_result_size(result); ++i) {
fprintf(stdout, "%s\t%f\n",
zinnia_result_value(result, i),
zinnia_result_score(result, i));
}
zinnia_result_destroy(result);
zinnia_character_destroy(character);
zinnia_recognizer_destroy(recognizer);
return 0;
}
% gcc example.c -lzinnia -o zinnia_sample % ./zinnia_sample
#include <iostream>
#include "zinnia.h"
int main(int argc, char **argv) {
zinnia::Recognizer *recognizer = zinnia::Recognizer::create();
if (!recognizer->open("/usr/local/lib/zinnia/model/tomoe/handwriting-ja.model")) {
std::cerr << recognizer->what() << std::endl;
return -1;
}
zinnia::Character *character = zinnia::Character::create();
character->clear();
character->set_width(300);
character->set_height(300);
character->add(0, 51, 29);
character->add(0, 117, 41);
character->add(1, 99, 65);
character->add(1, 219, 77);
character->add(2, 27, 131);
character->add(2, 261, 131);
character->add(3, 129, 17);
character->add(3, 57, 203);
character->add(4, 111, 71);
character->add(4, 219, 173);
character->add(5, 81, 161);
character->add(5, 93, 281);
character->add(6, 99, 167);
character->add(6, 207, 167);
character->add(6, 189, 245);
character->add(7, 99, 227);
character->add(7, 189, 227);
character->add(8, 111, 257);
character->add(8, 189, 245);
zinnia::Result *result = recognizer->classify(*character, 10);
if (!result) {
std::cerr << recognizer->what() << std::endl;
return -1;
}
for (size_t i = 0; i < result->size(); ++i) {
std::cout << result->value(i) << "\t" << result->score(i) << std::endl;
}
delete result;
delete character;
delete recognizer;
return 0;
}
Character class (zinnia_character_t / zinnia::Character) has method for an Sexp-based serialization.
(character (width canvas width) (height canvas height) (strokes ((0-th-stroke 0-th-strokey) ... (0-th-stroke 0-th-strokey)) ((1-th-stroke 0-th-strokey) ... (1-th-stroke 1-th-strokey)) ((2-th-stroke 2-th-strokey) ... (2-th-stroke 2-th-strokey)) ...))
Example
static const char sexp[] = "(character (width 1000)(height 1000)(strokes ((243 273)(393 450))((700 253)(343 486)(280 716)(393 866)(710 880))))"; zinnia_character_t *character = zinnia_character_new(); zinnia_character_parse(character, sexp);
static const char sexp[] = "(character (width 1000)(height 1000)(strokes ((243 273)(393 450))((700 253)(343 486)(280 716)(393 866)(710 880))))"; zinnia::Character *character = zinnia::Character::create(); character->parse(sexp);
In order to make a training with Zinnia, you need to make a golden (training) data. Training data is written in S-expression. Generally speaking, more data yields better quality. Zinnia only supports batch training so far.
Example
(character (value あ) (width 300) (height 300) (strokes ((54 58)(249 68)) ((147 10)(145 201)(182 252)) ((224 103)(149 230)(82 240)(53 204)(86 149)(182 139)(240 172)(248 224)(228 250)))) (character (value い) (width 300) (height 300) (strokes ((56 63)(43 213)(67 259)(94 243)) ((213 66)(231 171)(208 217)))) (character (value う) (width 300) (height 300) (strokes ((102 35)(187 45)) ((73 121)(167 105)(206 139)(198 211)(135 275)))) (character (value え) (width 300) (height 300) (strokes ((140 19)(162 38)) ((62 105)(208 100)(51 263)(128 205)(188 268)(260 252)))) (character (value お) (width 300) (height 300) (strokes ((64 94)(240 98)) ((148 35)(159 129)(140 199)(105 247)(64 228)(101 161)(192 161)(222 223)(189 257)) ((223 49)(253 89))))
You might want to look into *.s file in zinnia-tomoe as references
Use zinnia_learn to conduct training
% zinnia_learn train-file model-file ... learning: (8/3033) 8 .... learning: (9/3033) 9 ... learning: (10/3033) 々 ............ learning: (11/3033) 〆 .... learning: (12/3033) あ .... learning: (13/3033) い ....... learning: (14/3033) う ............... learning: (15/3033) え ...................................................................................... learning: (16/3033) お ...... learning: (17/3033) か ....... learning: (18/3033) き ....... learning: (19/3033) く .......
train-file is a training data and model-file is a model file zinnia_learn generates
use zinnia command for testing
% zinnia -m model-file < test-file
model-file is a file zinnia_learn generates. text-file is a test data written in Sexp format.
Once training completes successfully, model-file.txt (text file) and model-file (binary file) will be generated. text-model is an architecture-independent format, and model-file is an architecture-dependent binary model. Zinnia recognizer can only load binary model. zinnia_convert program converts text-model into binary model.
% zinnia_convert model-file.txt model-file
In addition, text-model can be converted into C-header file.
% zinnia_convert --name=my_model --make_header model-file.txt my_model.h
In order to use a header file, include my_model.h in C/C++ source file and load the data using zinnia_recognizer_open_from_ptr(). my_model(content of mode) and my_model_size(model size) is defined in the header file.
#include "my_model.h"
if (!zinnia_recognizer_open_from_ptr(recognizer, my_model, my_model_size)) {
/* error */
}
By omitting features having small score, model can be compressed with scarifying recognition accuracy.
% zinnia_convert -c 0.01 model-file.txt model1 % zinnia_convert -c 0.001 model-file.txt model2 % ls -h -l model* -rw-r--r-- 1 taku taku 3.1M 2008-07-14 12:50 model1 -rw-r--r-- 1 taku taku 25M 2008-07-14 12:50 model2
c is a threshold. Default setting 0.001. Larger threshold gives more aggressive compression.