diff --git a/README.md b/README.md index 6dc5bf9..9a6d9bc 100644 --- a/README.md +++ b/README.md @@ -11,6 +11,11 @@ ![](https://user-images.githubusercontent.com/952378/102350683-8a74db80-3fa5-11eb-8c7e-f743f7d6eae2.jpg) +

+ Input Image + Output Image +

+ ## Features * Document layout analysis using pixelwise segmentation models with support for 10 distinct segmentation classes: * background, [page border](https://ocr-d.de/en/gt-guidelines/trans/lyRand.html), [text region](https://ocr-d.de/en/gt-guidelines/trans/lytextregion.html#textregionen__textregion_), [text line](https://ocr-d.de/en/gt-guidelines/pagexml/pagecontent_xsd_Complex_Type_pc_TextLineType.html), [header](https://ocr-d.de/en/gt-guidelines/trans/lyUeberschrift.html), [image](https://ocr-d.de/en/gt-guidelines/trans/lyBildbereiche.html), [separator](https://ocr-d.de/en/gt-guidelines/trans/lySeparatoren.html), [marginalia](https://ocr-d.de/en/gt-guidelines/trans/lyMarginalie.html), [initial](https://ocr-d.de/en/gt-guidelines/trans/lyInitiale.html), [table](https://ocr-d.de/en/gt-guidelines/trans/lyTabellen.html) @@ -98,24 +103,35 @@ eynollah layout \ The following options can be used to further configure the processing: -| option | description | -|-------------------|:-------------------------------------------------------------------------------| -| `-fl` | full layout analysis including all steps and segmentation classes | -| `-light` | lighter and faster but simpler method for main region detection and deskewing | -| `-tll` | this indicates the light textline and should be passed with light version | -| `-tab` | apply table detection | -| `-ae` | apply enhancement (the resulting image is saved to the output directory) | -| `-as` | apply scaling | -| `-cl` | apply contour detection for curved text lines instead of bounding boxes | -| `-ib` | apply binarization (the resulting image is saved to the output directory) | -| `-ep` | enable plotting (MUST always be used with `-sl`, `-sd`, `-sa`, `-si` or `-ae`) | -| `-eoi` | extract only images to output directory (other processing will not be done) | -| `-ho` | ignore headers for reading order dectection | -| `-si ` | save image regions detected to this directory | -| `-sd ` | save deskewed image to this directory | -| `-sl ` | save layout prediction as plot to this directory | -| `-sp ` | save cropped page image to this directory | -| `-sa ` | save all (plot, enhanced/binary image, layout) to this directory | +| option | description | +|-------------------|:------------------------------------------------------------------------------- | +| `-fl` | full layout analysis including all steps and segmentation classes (recommended) | +| `-light` | lighter and faster but simpler method for main region detection and deskewing (recommended) | +| `-tll` | this indicates the light textline and should be passed with light version (recommended) | +| `-tab` | apply table detection | +| `-ae` | apply enhancement (the resulting image is saved to the output directory) | +| `-as` | apply scaling | +| `-cl` | apply contour detection for curved text lines instead of bounding boxes | +| `-ib` | apply binarization (the resulting image is saved to the output directory) | +| `-ep` | enable plotting (MUST always be used with `-sl`, `-sd`, `-sa`, `-si` or `-ae`) | +| `-eoi` | extract only images to output directory (other processing will not be done) | +| `-ho` | ignore headers for reading order dectection | +| `-si ` | save image regions detected to this directory | +| `-sd ` | save deskewed image to this directory | +| `-sl ` | save layout prediction as plot to this directory | +| `-sp ` | save cropped page image to this directory | +| `-sa ` | save all (plot, enhanced/binary image, layout) to this directory | +| `-thart` | threshold of artifical class in the case of textline detection. The default value is 0.1 | +| `-tharl` | threshold of artifical class in the case of layout detection. The default value is 0.1 | +| `-ocr` | do ocr | +| `-tr` | apply transformer ocr. Default model is a CNN-RNN model | +| `-bs_ocr` | ocr inference batch size. Default bs for trocr and cnn_rnn models are 2 and 8 respectively | +| `-ncu` | upper limit of columns in document image | +| `-ncl` | lower limit of columns in document image | +| `-slro` | skip layout detection and reading order | +| `-romb` | apply machine based reading order detection | +| `-ipe` | ignore page extraction | + If no further option is set, the tool performs layout detection of main regions (background, text, images, separators and marginals). @@ -133,7 +149,7 @@ The command-line interface for binarization can be called like this: eynollah binarization \ -i | -di \ -o \ - -m \ + -m ``` ### Image Enhancement @@ -141,7 +157,17 @@ TODO ### OCR -The OCR module performs text recognition using either CNN-RNN or TrOCR models. +

+ Input Image + Output Image +

+ +

+ Input Image + Output Image +

+ +The OCR module performs text recognition using either a CNN-RNN model or a Transformer model. The command-line interface for OCR can be called like this: @@ -150,9 +176,24 @@ eynollah ocr \ -i | -di \ -dx \ -o \ - -m | --model_name \ + -m | --model_name ``` +The following options can be used to further configure the ocr processing: + +| option | description | +|-------------------|:------------------------------------------------------------------------------- | +| `-dib` | directory of bins(files type must be '.png'). Prediction with both RGB and bins. | +| `-doit` | Directory containing output images rendered with the predicted text | +| `--model_name` | Specific model file path to use for OCR | +| `-trocr` | transformer ocr will be applied, otherwise cnn_rnn model | +| `-etit` | textlines images and text in xml will be exported into output dir (OCR training data) | +| `-nmtc` | cropped textline images will not be masked with textline contour | +| `-bs` | ocr inference batch size. Default bs for trocr and cnn_rnn models are 2 and 8 respectively | +| `-ds_pref` | add an abbrevation of dataset name to generated training data | +| `-min_conf` | minimum OCR confidence value. OCRs with textline conf lower than this will be ignored | + + ### Reading Order Detection Reading order detection can be performed either as part of layout analysis based on image input, or, currently under development, based on pre-existing layout analysis data in PAGE-XML format as input. diff --git a/docs/models.md b/docs/models.md index 3d296d5..b858630 100644 --- a/docs/models.md +++ b/docs/models.md @@ -18,7 +18,8 @@ Two Arabic/Persian terms form the name of the model suite: عين الله, whic See the flowchart below for the different stages and how they interact: -![](https://user-images.githubusercontent.com/952378/100619946-1936f680-331e-11eb-9297-6e8b4cab3c16.png) +eynollah_flowchart + ## Models @@ -151,15 +152,75 @@ This model is used for the task of illustration detection only. Model card: [Reading Order Detection]() -TODO +The model extracts the reading order of text regions from the layout by classifying pairwise relationships between them. A sorting algorithm then determines the overall reading sequence. + +### OCR + +We have trained three OCR models: two CNN-RNN–based models and one transformer-based TrOCR model. The CNN-RNN models are generally faster and provide better results in most cases, though their performance decreases with heavily degraded images. The TrOCR model, on the other hand, is computationally expensive and slower during inference, but it can possibly produce better results on strongly degraded images. + +#### CNN-RNN model: model_eynollah_ocr_cnnrnn_20250805 + +This model is trained on data where most of the samples are in Fraktur german script. + +| Dataset | Input | CER | WER | +|-----------------------|:-------|:-----------|:----------| +| OCR-D-GT-Archiveform | BIN | 0.02147 | 0.05685 | +| OCR-D-GT-Archiveform | RGB | 0.01636 | 0.06285 | + +#### CNN-RNN model: model_eynollah_ocr_cnnrnn_20250904 (Default) + +Compared to the model_eynollah_ocr_cnnrnn_20250805 model, this model is trained on a larger proportion of Antiqua data and achieves superior performance. + +| Dataset | Input | CER | WER | +|-----------------------|:------------|:-----------|:----------| +| OCR-D-GT-Archiveform | BIN | 0.01635 | 0.05410 | +| OCR-D-GT-Archiveform | RGB | 0.01471 | 0.05813 | +| BLN600 | RGB | 0.04409 | 0.08879 | +| BLN600 | Enhanced | 0.03599 | 0.06244 | + + +#### Transformer OCR model: model_eynollah_ocr_trocr_20250919 + +This transformer OCR model is trained on the same data as model_eynollah_ocr_trocr_20250919. + +| Dataset | Input | CER | WER | +|-----------------------|:------------|:-----------|:----------| +| OCR-D-GT-Archiveform | BIN | 0.01841 | 0.05589 | +| OCR-D-GT-Archiveform | RGB | 0.01552 | 0.06177 | +| BLN600 | RGB | 0.06347 | 0.13853 | + +##### Qualitative evaluation of the models + +| | | | | +|:---:|:---:|:---:|:---:| +| Image | cnnrnn_20250805 | cnnrnn_20250904 | trocr_20250919 | + + + +| | | | | +|:---:|:---:|:---:|:---:| +| Image | cnnrnn_20250805 | cnnrnn_20250904 | trocr_20250919 | + + +| | | | | +|:---:|:---:|:---:|:---:| +| Image | cnnrnn_20250805 | cnnrnn_20250904 | trocr_20250919 | + + +| | | | | +|:---:|:---:|:---:|:---:| +| Image | cnnrnn_20250805 | cnnrnn_20250904 | trocr_20250919 | + + ## Heuristic methods Additionally, some heuristic methods are employed to further improve the model predictions: * After border detection, the largest contour is determined by a bounding box, and the image cropped to these coordinates. -* For text region detection, the image is scaled up to make it easier for the model to detect background space between text regions. +* Unlike the non-light version, where the image is scaled up to help the model better detect the background spaces between text regions, the light version uses down-scaled images. In this case, introducing an artificial class along the boundaries of text regions and text lines has helped to isolate and separate the text regions more effectively. * A minimum area is defined for text regions in relation to the overall image dimensions, so that very small regions that are noise can be filtered out. -* Deskewing is applied on the text region level (due to regions having different degrees of skew) in order to improve the textline segmentation result. -* After deskewing, a calculation of the pixel distribution on the X-axis allows the separation of textlines (foreground) and background pixels. -* Finally, using the derived coordinates, bounding boxes are determined for each textline. +* In the non-light version, deskewing is applied at the text-region level (since regions may have different degrees of skew) to improve text-line segmentation results. In contrast, the light version performs deskewing only at the page level to enhance margin detection and heuristic reading-order estimation. +* After deskewing, a calculation of the pixel distribution on the X-axis allows the separation of textlines (foreground) and background pixels (only in non-light version). +* Finally, using the derived coordinates, bounding boxes are determined for each textline (only in non-light version). +* As mentioned above, the reading order can be determined using a model; however, this approach is computationally expensive, time-consuming, and less accurate due to the limited amount of ground-truth data available for training. Therefore, our tool uses a heuristic reading-order detection method as the default. The heuristic approach relies on headers and separators to determine the reading order of text regions.