English | [简体中文](../doc_ch/PP-OCRv3_introduction.md)
# PP-OCRv3
- [1. Introduction](#1)
- [2. Optimization for Text Detection Model](#2)
- [3. Optimization for Text Recognition Model](#3)
- [4. End-to-end Evaluation](#4)
## 1. Introduction
PP-OCRv3 is further upgraded on the basis of PP-OCRv2. The overall framework of PP-OCRv3 is same as that of PP-OCRv2. The text detection model and text recognition model are further optimized, respectively. Specifically, the detection network is still optimized based on DBNet, and base model of recognition network is replaced from CRNN to [SVTR](https://arxiv.org/abs/2205.00159, which is recorded in IJCAI 2022. The block diagram of the PP-OCRv3 system is as follows (the pink box is the new policies of PP-OCRv3):
There are 9 optimization tricks for text detection and recognition models in PP-OCRv3, which is as follows.
- Text detection:
- LK-PAN: PAN structure with large receptive field;
- DML: mutual learning strategy for teacher model;
- RSE-FPN: FPN structure of residual attention mechanism;
- Text recognition:
- SVTR_LCNet: Light-weight text recognition network;
- GTC: training strategy using Attention to guide CTC;
- TextConAug: A data augmentation strategy for mining textual context information;
- TextRotNet: self-supervised strategy to optimize the pretrained model;
- UDML: unified deep mutual learning strategy;
- UIM: unlabeled data mining strategy.
From the effect point of view, when the speed is comparable, the accuracy of various scenes has been greatly improved:
Finally, in the case of comparable inference speed, PP-OCRv3 significantly outperforms PP-OCRv2 in terms of accuracy in multiple scenarios.
- In Chinese scenarios, PP-OCRv3 outperforms PP-OCRv2 by more than 5%.
- In English scenarios, PP-OCRv3 outperforms PP-OCRv2 by more than 11%.
- In multi-language scenarios, more than 80 languages and corresponding models are optimized, the average accuracy is increased by over 5%.
## 2. Optimization for Text Detection Model
The PP-OCRv3 detection model is an upgrade of the [CML](https://arxiv.org/pdf/2109.03144.pdf) (Collaborative Mutual Learning) text detection distillation strategy in PP-OCRv2. As shown in the figure below, the core idea of CML combines ① the distillation of the traditional Teacher to guide Students and ② the DML mutual learning between the Students network, which allows the Students network to learn from each other. PP-OCRv3 further optimizes the effect of teacher model and student model respectively. For the Teacher model,
we propose a PAN structure LKPAN with a larger receptive field and use the DML (Deep Mutual Learning) distillation strategy to optimizing the Teacher model. For the student model,
we propose a lightweight FPN structure RSEFPN to improve the accuracy of the student model.
The ablation experiment is as follows:
|ID|Strategy|Model Size|Hmean|The Inference Time（cpu + mkldnn)|
|-|-|-|-|-|
|baseline teacher|PP-OCR server|49M|83.2%|171ms|
|teacher1|DB-R50-LK-PAN|124M|85.0%|396ms|
|teacher2|DB-R50-LK-PAN-DML|124M|86.0%|396ms|
|baseline student|PP-OCRv2|3M|83.2%|117ms|
|student0|DB-MV3-RSE-FPN|3.6M|84.5%|124ms|
|student1|DB-MV3-CML（teacher2）|3M|84.3%|117ms|
|student2|DB-MV3-RSE-FPN-CML（teacher2）|3.6M|85.4%|124ms|
The environment: Intel Gold 6148 CPU, with MKLDNN acceleration enabled during inference.
**(1) LK-PAN: PAN structure with large receptive field**
LK-PAN (Large Kernel PAN) is a lightweight [PAN](https://arxiv.org/pdf/1803.01534.pdf) structure with a larger receptive field. The core is to change the convolution kernel in the path augmentation of the PAN structure from `3*3` to `9*9`. By increasing the convolution kernel, the receptive field covered by each position of the feature map is improved, and it is easier to detect text in large fonts and text with extreme aspect ratios. Using the LK-PAN, the hmean of the teacher model can be improved from 83.2% to 85.0%.
**(2) DML: The Mutual Learning Strategy for Teacher Model**
The [DML](https://arxiv.org/abs/1706.00384) method, as shown in the figure below, can effectively improve the accuracy of the text detection model by learning from each other with two models with the same structure. The teacher model adopts the DML strategy, and the hmean is increased from 85% to 86%. By updating the teacher model of CML in PP-OCRv2 to the above-mentioned higher-precision teacher model, the hmean of the student model can be further improved from 83.2% to 84.3%.
**(3) RSE-FPN: FPN structure of residual attention mechanism**
RSE-FPN (Residual Squeeze-and-Excitation FPN) is shown in the figure below. RSEFPN introduces the residual structure and the channel attention structure, and replaces the convolutional layer in the FPN with the RSEConv layer of the channel attention structure to improve the representation ability of the feature map.
Considering that the number of FPN channels in the detection model of PP-OCRv2 is very small, only 96, if SEblock is directly used to replace the convolution in FPN, the features of some channels will be suppressed, and the accuracy will be reduced. The introduction of residual structure in RSEConv will alleviate the above problems and improve the text detection effect. By further updating the FPN structure of the student model of CML in PP-OCRv2 to RSE-FPN, the hmean of the student model can be further improved from 84.3% to 85.4%.
## 3. Optimization for Text Recognition Model
The recognition module of PP-OCRv3 is optimized based on the text recognition algorithm [SVTR] (https://arxiv.org/abs/2205.00159). RNN is abandoned in SVTR, and the context information of the text line image is more effectively mined by introducing the Transformers structure, thereby improving the text recognition ability. The recognition model of PP-OCRv2 was directly replaced with SVTR_Tiny, and the recognition accuracy increased from 74.8% to 80.1% (+5.3%), but the prediction speed was nearly 11 times slower, and it took nearly 100ms to predict a text line on the CPU. Therefore, as shown in the figure below, PP-OCRv3 adopts the following six optimization strategies to accelerate the recognition model.
Based on the above strategy, compared with PP-OCRv2, the PP-OCRv3 recognition model further improves the accuracy by 4.6% with comparable speed. The specific ablation experiments are as follows:
| ID | strategy | Model size | accuracy | prediction speed（CPU + MKLDNN)|
|-----|-----|--------|----| --- |
| 01 | PP-OCRv2 | 8M | 74.8% | 8.54ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 03 | SVTR_LCNet(h32) | 12M | 71.9% | 6.6ms |
| 04 | SVTR_LCNet(h48) | 12M | 73.98% | 7.6ms |
| 05 | + GTC | 12M | 75.8% | 7.6ms |
| 06 | + TextConAug | 12M | 76.3% | 7.6ms |
| 07 | + TextRotNet | 12M | 76.9% | 7.6ms |
| 08 | + UDML | 12M | 78.4% | 7.6ms |
| 09 | + UIM | 12M | 79.4% | 7.6ms |
Note: When testing the speed, the input image shape of Experiment 01-03 is (3, 32, 320), and the input image shape of 04-08 is (3, 48, 320). In the actual prediction, the image is a variable-length input, and the speed will vary. Test environment: Intel Gold 6148 CPU, with MKLDNN acceleration enabled during prediction.
**（1）SVTR_LCNet：Lightweight Text Recognition Network**
SVTR_LCNet is for text recognition tasks, which is a lightweight text recognition network fused by Transformer-based [SVTR](https://arxiv.org/abs/2205.00159) network and lightweight CNN network [PP-LCNet](https://arxiv.org/abs/ 2109.15099). Using this network, the prediction speed is 20% better than the recognition model of PP-OCRv2, but because the distillation strategy is not adopted, the effect of the recognition model is slightly worse. In addition, the height of the input image is further increased from 32 to 48, and the prediction speed is slightly slower, but the model effect is greatly improved, and the recognition accuracy reaches 73.98% (+2.08%), which is close to the recognition model effect of PP-OCRv2 using the distillation strategy.
SVTR_Tiny network structure is as follows：
Due to the limited model structure supported by the MKLDNN acceleration library, SVTR is 10 times slower than PP-OCRv2 on CPU+MKLDNN. PP-OCRv3 expects to improve the accuracy of the model without bringing additional inference time. Through analysis, it is found that the main time-consuming module of SVTR_Tiny structure is Mixing Block, so we have carried out a series of optimizations to the structure of SVTR_Tiny (for detailed speed data, please refer to the ablation experiment table below):
1. Replace the first half of the SVTR network with the first three stages of PP-LCNet, retain 4 Global Mixing Blocks, the accuracy is 76%, and the speedup is 69%. The network structure is as follows:
2. Reduce the number of Global Mixing Blocks from 4 to 2, the accuracy is 72.9%, and the speedup is 69%. The network structure is as follows:
3. The experiment found that the prediction speed of the Global Mixing Block is related to the shape of the input features. Therefore, after moving the position of the Global Mixing Block to the back of pooling layer, the accuracy dropped to 71.9%, and the speed surpassed the PP-OCRv2-baseline based on the CNN structure by 22%. The network structure is as follows:
The specific ablation experiments are as follows:
| ID | strategy | Model size | accuracy | prediction speed（CPU + MKLDNN)|
|-----|-----|--------|----| --- |
| 01 | PP-OCRv2-baseline | 8M | 69.3% | 8.54ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 03 | SVTR_LCNet(G4) | 9.2M | 76% | 30ms |
| 04 | SVTR_LCNet(G2) | 13M | 72.98% | 9.37ms |
| 05 | SVTR_LCNet(h32) | 12M | 71.9% | 6.6ms |
| 06 | SVTR_LCNet(h48) | 12M | 73.98% | 7.6ms |
Note: When testing the speed, the input image shape of 01-05 are all (3, 32, 320); PP-OCRv2-baseline represents the model trained without distillation method
**（2）GTC：Attention guides CTC training strategy**
[GTC](https://arxiv.org/pdf/2002.01276.pdf)（Guided Training of CTC），using the Attention module and loss to guide the CTC loss training and fuse the expression of multiple text features is an effective strategy to improve text recognition. Using this strategy, the Attention module is completely removed during prediction, and no time-consuming is added in the inference stage, and the accuracy of the recognition model is further improved to 75.8% (+1.82%). The training process is as follows:
**（3）TextConAug：Data Augmentation Strategy for Mining Text Context Information**
TextConAug is a data augmentation strategy for mining textual context information. The main idea comes from the paper [ConCLR](https://www.cse.cuhk.edu.hk/~byu/papers/C139-AAAI2022-ConCLR.pdf) , the author proposes ConAug data augmentation to connect 2 different images in a batch to form new images and perform self-supervised comparative learning. PP-OCRv3 applies this method to supervised learning tasks, and designs the TextConAug data augmentation method, which can enrich the context information of training data and improve the diversity of training data. Using this strategy, the accuracy of the recognition model is further improved to 76.3% (+0.5%). The schematic diagram of TextConAug is as follows:
**（4）TextRotNet：Self-Supervised Pretrained Models**
TextRotNet is a pre-training model, which is trained by using a large amount of unlabeled text line data in a self-supervised manner. Refer to the paper [STR-Fewer-Labels](https://github.com/ku21fan/STR-Fewer-Labels). This model can initialize the initial weights of SVTR_LCNet, which helps the text recognition model to converge to a better position. Using this strategy, the accuracy of the recognition model is further improved to 76.9% (+0.6%). The TextRotNet training process is shown in the following figure:
**（5）UDML：Unified-Deep Mutual Learning**
UDML (Unified-Deep Mutual Learning) is a strategy adopted in PP-OCRv2 that is very effective for text recognition to improve the model effect. In PP-OCRv3, for two different SVTR_LCNet and Attention structures, the feature map of PP-LCNet, the output of the SVTR module and the output of the Attention module between them are simultaneously supervised and trained. Using this strategy, the accuracy of the recognition model is further improved to 78.4% (+1.5%).
**（6）UIM：Unlabeled Images Mining**
UIM (Unlabeled Images Mining) is a very simple unlabeled data mining scheme. The core idea is to use a high-precision text recognition model to predict unlabeled data, obtain pseudo-labels, and select samples with high prediction confidence as training data for training small models. Using this strategy, the accuracy of the recognition model is further improved to 79.4% (+1%).
## 4. End-to-end Evaluation
After the optimization strategies mentioned above, under he condition of comparable speed, PP-OCRv3 outperforms PP-OCRv2 by 5% in terms of end-to-end Hmean for Chinese scenarios. The specific metrics are shown as follows.
| Model | Hmean | Model Size (M) | Time Cost (CPU, ms) | Time Cost (T4 GPU, ms) |
|-----|-----|--------|----| --- |
| PP-OCR mobile | 50.3% | 8.1 | 356 | 116 |
| PP-OCR server | 57.0% | 155.1 | 1056 | 200 |
| PP-OCRv2 | 57.6% | 11.6 | 330 | 111 |
| PP-OCRv3 | 62.9% | 15.6 | 331 | 86.64 |
Test environment:
- CPU: Intel Gold 6148, and MKLDNN acceleration is enabled during CPU inference.
In addition to updating the recognition model for Chinese, the recognition model for English is also optimized with an increasement of 11% for end-to-end Hmean, which is shown as follows.
| Model | Recall | Precision | Hmean |
|-----|-----|--------|----|
| PP-OCR_en | 38.99% | 45.91% | 42.17% |
| PP-OCRv3_en | 50.95% | 55.53% | 53.14% |
At the same time, more than 80 language recognition models that have been supported have been upgraded this time, and the recognition accuracy of the four language families with evaluation sets has increased by more than 5% on average, which is shown as follows.
| Model | Latin | Arabic | Japanese | Korean |
|-----|-----|--------|----| --- |
| PP-OCR_mul | 69.6% | 40.5% | 38.5% | 55.4% |
| PP-OCRv3_mul | 75.2% | 45.37% | 45.8% | 60.1% |
