ABBYY FineReader SDK如何定义最小识别率

Question

有人使用fineReader ABBYYSDK10吗？我很好奇，在图像ocr处理之后，是否有可能获得数据挖掘的成功率。

对于这样的场景，我们有从图像收集数据的工作流程，如果识别结果低于90%，那么我们将我们的批次进行可视化验证/更正。

对于软件开发工具包的处理，我使用的是.net --这不是很重要，但是……以防万一

我怎样才能达到这个数字呢？谢谢你的建议

Answer 1

没有“全局识别置信度”属性。开发人员需要使用他们自己的置信度标准来计算它。最简单的方法是遍历每个字符，检查CharParams.IsSuspicious属性。以下是FREngine 11的代码示例(C#

    //Statistics counters 

    //Count of all suspicious symbols in layout
    private int suspiciousSymbolsCount;
    //Count of all unrecognized symbols in layout
    private int unrecognizedSymbolsCount;
    //Count of all nonspace symbols in layout
    private int allSymbolsCount;
    //Count of all words in layout
    private int allWordsCount;
    //Count of all not dictionary word in layout
    private int notDictionaryWordsCount;
    private void processImage()
    {
        // Create document
        FRDocument document = engineLoader.Engine.CreateFRDocument();

        try {
            // Add image file to document
            displayMessage( "Loading image..." );
            string imagePath = Path.Combine( FreConfig.GetSamplesFolder(), @"SampleImages\Demo.tif" );

            document.AddImageFile( imagePath, null, null );

            //Recognize document
            displayMessage( "Recognizing..." );
            document.Process( null );

            // Calculate text statistics
            displayMessage( "Calculating statistics..." );
            clearStatistics();
            for( int i = 0; i < document.Pages.Count; i++ ) {
                calculateStatisticsForLayout( document.Pages[i].Layout );
            }

            //show calculated statistics
            displayStatistics();

        } catch( Exception error ) {
            MessageBox.Show( this, error.Message, this.Text, MessageBoxButtons.OK, MessageBoxIcon.Error );
        }
        finally {
            // Close document
            document.Close();
        }
    }
    private void calculateStatisticsForLayout( Layout layout )
    {    
        LayoutBlocks blocks = layout.Blocks;
        for( int index = 0; index < blocks.Count; index++ ) {
            calculateStatisticsForBlock( blocks[index] );
        }
    }

    void calculateStatisticsForBlock( IBlock block )
    {           
        if( block.Type == BlockTypeEnum.BT_Text ) {
            calculateStatisticsForTextBlock( block.GetAsTextBlock() );
        } else if( block.Type == BlockTypeEnum.BT_Table ) {
            calculateStatisticsForTableBlock( block.GetAsTableBlock() );
        }
    }

    void calculateStatisticsForTextBlock( TextBlock textBlockProperties )
    {
        calculateStatisticsForText( textBlockProperties.Text );
    }

    void calculateStatisticsForTableBlock( TableBlock tableBlockProperties )
    {
        for( int index = 0; index < tableBlockProperties.Cells.Count; index++ ) {
            calculateStatisticsForBlock( tableBlockProperties.Cells[index].Block );
        }
    }

    void calculateStatisticsForText( Text text ) 
    {
        Paragraphs paragraphs = text.Paragraphs;
        for( int index = 0; index < paragraphs.Count; index++ ) {
            calculateStatisticsForParagraph( paragraphs[index] );
        }
    }

    void calculateStatisticsForParagraph( Paragraph paragraph )
    {
        calculateCharStatisticsForParagraph( paragraph );

        calculateWordStatisticsForParagraph( paragraph );
    }

    void calculateCharStatisticsForParagraph( Paragraph paragraph )
    {
        for( int index = 0; index < paragraph.Text.Length; index++ )
        {
            calculateStatisticsForChar( paragraph, index );
        }
    }

    void calculateStatisticsForChar( Paragraph paragraph, int charIndex )
    {
        CharParams charParams = engineLoader.Engine.CreateCharParams();
        paragraph.GetCharParams( charIndex, charParams );
        if( charParams.IsSuspicious ) 
        {
            suspiciousSymbolsCount++;
        }

        if( isUnrecognizedSymbol( paragraph.Text[charIndex] ) ) 
        {
            unrecognizedSymbolsCount++;
        }

        if( paragraph.Text[charIndex] != ' ' ) 
        {
            allSymbolsCount++;
        }
    }

    void calculateWordStatisticsForParagraph( Paragraph paragraph )
    {
        allWordsCount += paragraph.Words.Count;

        for( int index = 0; index < paragraph.Words.Count; index++ ) 
        {
            if( !paragraph.Words[index].IsWordFromDictionary ) 
            {
                notDictionaryWordsCount ++;
            }
        }
    }

    bool isUnrecognizedSymbol( char symbol )
    {
        //it is special constant used by FREngine recogniser
        return ( symbol == 0x005E );
    }

    void displayStatistics()
    {
        labelAllSymbols.Text = "All symbols: " + allSymbolsCount.ToString();
        labelSuspiciosSymbols.Text = "Suspicious symbols: " + suspiciousSymbolsCount.ToString();
        labelUnrecognizedSymbols.Text = "Unrecognized symbols: " + unrecognizedSymbolsCount.ToString();

        labelAllWords.Text = "All words: " + allWordsCount.ToString();
        labelNotDictionaryWords.Text = "Non-dictionary words: " + notDictionaryWordsCount.ToString();
    }

Answer 2

IMHO没有这样的“全局置信度”值--但你可以很容易地通过取每个字符的置信度并求出总数的平均值来获得这个值。然而，我认为你应该将你的请求直接发送到ABBYY的论坛或支持电子邮件地址，看看他们的建议是什么。

如果我使用这个引擎，我真的不可能告诉你我会得到什么样的置信度，因为这一切都依赖于图像的质量，字体的大小等等:业界没有‘平均文档’这样的东西来作为他们数据的基础。

祝好运!

Answer 3

FRE SDK识别的结果仅包含文本块或表块中的文本。我建议你有一个全局字数统计变量。

1. 运行异步方法来遍历单词并获取单词中的可疑字符数。(IsSuspicious)
2. Find每页characters
3. (words有可疑字符的总字数)/(总字数)乘以100。

2/4等于0.5。乘以0.5 * 100 = 50%。这是用于检查可疑字符的accuracy.The代码样本，上面来自abbyy的另一个答案给出了置信度。