Added new LSTM-based neural network line recognizer

Author: theraysmith

Makefile.am

@@ -16,7 +16,7 @@ endif
 
 .PHONY: install-langs ScrollView.jar install-jars training
 
-SUBDIRS = ccutil viewer cutil opencl ccstruct dict classify wordrec textord
+SUBDIRS = arch ccutil viewer cutil opencl ccstruct dict classify wordrec textord lstm
 if !NO_CUBE_BUILD
     SUBDIRS += neural_networks/runtime cube
 endif

api/Makefile.am

@@ -1,5 +1,6 @@
 AM_CPPFLAGS += -DLOCALEDIR=\"$(localedir)\"\
     -DUSE_STD_NAMESPACE \
+    -I$(top_srcdir)/arch -I$(top_srcdir)/lstm \
     -I$(top_srcdir)/ccutil -I$(top_srcdir)/ccstruct -I$(top_srcdir)/cube \
     -I$(top_srcdir)/viewer \
     -I$(top_srcdir)/textord -I$(top_srcdir)/dict \
@@ -27,6 +28,9 @@ libtesseract_api_la_LIBADD = \
     ../wordrec/libtesseract_wordrec.la \
     ../classify/libtesseract_classify.la \
     ../dict/libtesseract_dict.la \
+    ../arch/libtesseract_avx.la \
+    ../arch/libtesseract_sse.la \
+    ../lstm/libtesseract_lstm.la \
     ../ccstruct/libtesseract_ccstruct.la \
     ../cutil/libtesseract_cutil.la \
     ../viewer/libtesseract_viewer.la \
@@ -57,6 +61,9 @@ libtesseract_la_LIBADD = \
     ../wordrec/libtesseract_wordrec.la \
     ../classify/libtesseract_classify.la \
     ../dict/libtesseract_dict.la \
+    ../arch/libtesseract_avx.la \
+    ../arch/libtesseract_sse.la \
+    ../lstm/libtesseract_lstm.la \
     ../ccstruct/libtesseract_ccstruct.la \
     ../cutil/libtesseract_cutil.la \
     ../viewer/libtesseract_viewer.la \

api/baseapi.cpp

@@ -121,7 +121,6 @@ TessBaseAPI::TessBaseAPI()
     block_list_(NULL),
     page_res_(NULL),
     input_file_(NULL),
-    input_image_(NULL),
     output_file_(NULL),
     datapath_(NULL),
     language_(NULL),
@@ -515,19 +514,19 @@ void TessBaseAPI::ClearAdaptiveClassifier() {
 
 /**
  * Provide an image for Tesseract to recognize. Format is as
- * TesseractRect above. Does not copy the image buffer, or take
- * ownership. The source image may be destroyed after Recognize is called,
- * either explicitly or implicitly via one of the Get*Text functions.
+ * TesseractRect above. Copies the image buffer and converts to Pix.
  * SetImage clears all recognition results, and sets the rectangle to the
  * full image, so it may be followed immediately by a GetUTF8Text, and it
  * will automatically perform recognition.
  */
 void TessBaseAPI::SetImage(const unsigned char* imagedata,
                            int width, int height,
                            int bytes_per_pixel, int bytes_per_line) {
-  if (InternalSetImage())
+  if (InternalSetImage()) {
     thresholder_->SetImage(imagedata, width, height,
                            bytes_per_pixel, bytes_per_line);
+    SetInputImage(thresholder_->GetPixRect());
+  }
 }
 
 void TessBaseAPI::SetSourceResolution(int ppi) {
@@ -539,18 +538,17 @@ void TessBaseAPI::SetSourceResolution(int ppi) {
 
 /**
  * Provide an image for Tesseract to recognize. As with SetImage above,
- * Tesseract doesn't take a copy or ownership or pixDestroy the image, so
- * it must persist until after Recognize.
+ * Tesseract takes its own copy of the image, so it need not persist until
+ * after Recognize.
  * Pix vs raw, which to use?
- * Use Pix where possible. A future version of Tesseract may choose to use Pix
- * as its internal representation and discard IMAGE altogether.
- * Because of that, an implementation that sources and targets Pix may end up
- * with less copies than an implementation that does not.
+ * Use Pix where possible. Tesseract uses Pix as its internal representation
+ * and it is therefore more efficient to provide a Pix directly.
  */
 void TessBaseAPI::SetImage(Pix* pix) {
-  if (InternalSetImage())
+  if (InternalSetImage()) {
     thresholder_->SetImage(pix);
-  SetInputImage(pix);
+    SetInputImage(thresholder_->GetPixRect());
+  }
 }
 
 /**
@@ -693,8 +691,8 @@ Boxa* TessBaseAPI::GetComponentImages(PageIteratorLevel level,
       if (pixa != NULL) {
         Pix* pix = NULL;
         if (raw_image) {
-          pix = page_it->GetImage(level, raw_padding, input_image_,
-                                  &left, &top);
+          pix = page_it->GetImage(level, raw_padding, GetInputImage(), &left,
+                                  &top);
         } else {
           pix = page_it->GetBinaryImage(level);
         }
@@ -849,13 +847,17 @@ int TessBaseAPI::Recognize(ETEXT_DESC* monitor) {
   } else if (tesseract_->tessedit_resegment_from_boxes) {
     page_res_ = tesseract_->ApplyBoxes(*input_file_, false, block_list_);
   } else {
-    // TODO(rays) LSTM here.
-    page_res_ = new PAGE_RES(false,
+    page_res_ = new PAGE_RES(tesseract_->AnyLSTMLang(),
                              block_list_, &tesseract_->prev_word_best_choice_);
   }
   if (page_res_ == NULL) {
     return -1;
   }
+  if (tesseract_->tessedit_train_line_recognizer) {
+    tesseract_->TrainLineRecognizer(*input_file_, *output_file_, block_list_);
+    tesseract_->CorrectClassifyWords(page_res_);
+    return 0;
+  }
   if (tesseract_->tessedit_make_boxes_from_boxes) {
     tesseract_->CorrectClassifyWords(page_res_);
     return 0;
@@ -938,17 +940,10 @@ int TessBaseAPI::RecognizeForChopTest(ETEXT_DESC* monitor) {
   return 0;
 }
 
-void TessBaseAPI::SetInputImage(Pix *pix) {
-  if (input_image_)
-    pixDestroy(&input_image_);
-  input_image_ = NULL;
-  if (pix)
-    input_image_ = pixCopy(NULL, pix);
-}
+// Takes ownership of the input pix.
+void TessBaseAPI::SetInputImage(Pix* pix) { tesseract_->set_pix_original(pix); }
 
-Pix* TessBaseAPI::GetInputImage() {
-  return input_image_;
-}
+Pix* TessBaseAPI::GetInputImage() { return tesseract_->pix_original(); }
 
 const char * TessBaseAPI::GetInputName() {
   if (input_file_)
@@ -992,8 +987,7 @@ bool TessBaseAPI::ProcessPagesFileList(FILE *flist,
   }
 
   // Begin producing output
-  const char* kUnknownTitle = "";
-  if (renderer && !renderer->BeginDocument(kUnknownTitle)) {
+  if (renderer && !renderer->BeginDocument(unknown_title_)) {
     return false;
   }
 
@@ -1105,7 +1099,6 @@ bool TessBaseAPI::ProcessPagesInternal(const char* filename,
                                        const char* retry_config,
                                        int timeout_millisec,
                                        TessResultRenderer* renderer) {
-#ifndef ANDROID_BUILD
   PERF_COUNT_START("ProcessPages")
   bool stdInput = !strcmp(filename, "stdin") || !strcmp(filename, "-");
   if (stdInput) {
@@ -1162,8 +1155,7 @@ bool TessBaseAPI::ProcessPagesInternal(const char* filename,
   }
 
   // Begin the output
-  const char* kUnknownTitle = "";
-  if (renderer && !renderer->BeginDocument(kUnknownTitle)) {
+  if (renderer && !renderer->BeginDocument(unknown_title_)) {
     pixDestroy(&pix);
     return false;
   }
@@ -1185,9 +1177,6 @@ bool TessBaseAPI::ProcessPagesInternal(const char* filename,
   }
   PERF_COUNT_END
   return true;
-#else
-  return false;
-#endif
 }
 
 bool TessBaseAPI::ProcessPage(Pix* pix, int page_index, const char* filename,
@@ -2107,10 +2096,6 @@ void TessBaseAPI::End() {
     delete input_file_;
     input_file_ = NULL;
   }
-  if (input_image_ != NULL) {
-    pixDestroy(&input_image_);
-    input_image_ = NULL;
-  }
   if (output_file_ != NULL) {
     delete output_file_;
     output_file_ = NULL;

api/baseapi.h

@@ -20,8 +20,8 @@
 #ifndef TESSERACT_API_BASEAPI_H__
 #define TESSERACT_API_BASEAPI_H__
 
-#define TESSERACT_VERSION_STR "3.05.00dev"
-#define TESSERACT_VERSION 0x030500
+#define TESSERACT_VERSION_STR "4.00.00alpha"
+#define TESSERACT_VERSION 0x040000
 #define MAKE_VERSION(major, minor, patch) (((major) << 16) | ((minor) << 8) | \
                                             (patch))
 
@@ -142,6 +142,7 @@ class TESS_API TessBaseAPI {
    * is stored in the PDF so we need that as well.
    */
   const char* GetInputName();
+  // Takes ownership of the input pix.
   void SetInputImage(Pix *pix);
   Pix* GetInputImage();
   int GetSourceYResolution();
@@ -333,9 +334,7 @@ class TESS_API TessBaseAPI {
 
   /**
    * Provide an image for Tesseract to recognize. Format is as
-   * TesseractRect above. Does not copy the image buffer, or take
-   * ownership. The source image may be destroyed after Recognize is called,
-   * either explicitly or implicitly via one of the Get*Text functions.
+   * TesseractRect above. Copies the image buffer and converts to Pix.
    * SetImage clears all recognition results, and sets the rectangle to the
    * full image, so it may be followed immediately by a GetUTF8Text, and it
    * will automatically perform recognition.
@@ -345,13 +344,11 @@ class TESS_API TessBaseAPI {
 
   /**
    * Provide an image for Tesseract to recognize. As with SetImage above,
-   * Tesseract doesn't take a copy or ownership or pixDestroy the image, so
-   * it must persist until after Recognize.
+   * Tesseract takes its own copy of the image, so it need not persist until
+   * after Recognize.
    * Pix vs raw, which to use?
-   * Use Pix where possible. A future version of Tesseract may choose to use Pix
-   * as its internal representation and discard IMAGE altogether.
-   * Because of that, an implementation that sources and targets Pix may end up
-   * with less copies than an implementation that does not.
+   * Use Pix where possible. Tesseract uses Pix as its internal representation
+   * and it is therefore more efficient to provide a Pix directly.
    */
   void SetImage(Pix* pix);
 
@@ -866,7 +863,6 @@ class TESS_API TessBaseAPI {
   BLOCK_LIST*       block_list_;      ///< The page layout.
   PAGE_RES*         page_res_;        ///< The page-level data.
   STRING*           input_file_;      ///< Name used by training code.
-  Pix*              input_image_;     ///< Image used for searchable PDF
   STRING*           output_file_;     ///< Name used by debug code.
   STRING*           datapath_;        ///< Current location of tessdata.
   STRING*           language_;        ///< Last initialized language.
@@ -902,6 +898,12 @@ class TESS_API TessBaseAPI {
                                  int timeout_millisec,
                                  TessResultRenderer* renderer,
                                  int tessedit_page_number);
+  // There's currently no way to pass a document title from the
+  // Tesseract command line, and we have multiple places that choose
+  // to set the title to an empty string. Using a single named
+  // variable will hopefully reduce confusion if the situation changes
+  // in the future.
+  const char *unknown_title_ = "";
 };  // class TessBaseAPI.
 
 /** Escape a char string - remove &<>"' with HTML codes. */

api/pdfrenderer.cpp

@@ -620,7 +620,6 @@ bool TessPDFRenderer::BeginDocumentHandler() {
   AppendPDFObject(buf);
 
   // FONT DESCRIPTOR
-  const int kCharHeight = 2;  // Effect: highlights are half height
   n = snprintf(buf, sizeof(buf),
                "7 0 obj\n"
                "<<\n"
@@ -636,10 +635,10 @@ bool TessPDFRenderer::BeginDocumentHandler() {
                "  /Type /FontDescriptor\n"
                ">>\n"
                "endobj\n",
-               1000 / kCharHeight,
-               1000 / kCharHeight,
+               1000,
+               1000,
                1000 / kCharWidth,
-               1000 / kCharHeight,
+               1000,
                8L      // Font data
                );
   if (n >= sizeof(buf)) return false;

api/renderer.h

@@ -77,7 +77,7 @@ class TESS_API TessResultRenderer {
     bool EndDocument();
 
     const char* file_extension() const { return file_extension_; }
-    const char* title() const { return title_; }
+    const char* title() const { return title_.c_str(); }
 
     /**
      * Returns the index of the last image given to AddImage
@@ -126,7 +126,7 @@ class TESS_API TessResultRenderer {
 
   private:
     const char* file_extension_;  // standard extension for generated output
-    const char* title_;           // title of document being renderered
+    STRING title_;                // title of document being renderered
     int imagenum_;                // index of last image added
 
     FILE* fout_;                  // output file pointer

arch/Makefile.am

@@ -0,0 +1,29 @@
+AM_CPPFLAGS += -I$(top_srcdir)/ccutil
+AUTOMAKE_OPTIONS = subdir-objects
+SUBDIRS =
+AM_CXXFLAGS =
+
+if VISIBILITY
+AM_CXXFLAGS += -fvisibility=hidden -fvisibility-inlines-hidden
+AM_CPPFLAGS += -DTESS_EXPORTS
+endif
+
+include_HEADERS = \
+	dotproductavx.h dotproductsse.h
+
+noinst_HEADERS = 
+
+if !USING_MULTIPLELIBS
+noinst_LTLIBRARIES = libtesseract_avx.la libtesseract_sse.la
+else
+lib_LTLIBRARIES = libtesseract_avx.la libtesseract_sse.la
+libtesseract_avx_la_LDFLAGS = -version-info $(GENERIC_LIBRARY_VERSION)
+libtesseract_sse_la_LDFLAGS = -version-info $(GENERIC_LIBRARY_VERSION)
+endif
+libtesseract_avx_la_CXXFLAGS = -mavx
+libtesseract_sse_la_CXXFLAGS = -msse4.1
+
+libtesseract_avx_la_SOURCES = dotproductavx.cpp
+
+libtesseract_sse_la_SOURCES = dotproductsse.cpp
+