Image segmentation, the process of extracting specific regions from an image, is applicable in various fields such as autonomous driving, medical image analysis, and augmented reality. A prominent method for image segmentation is machine learning-based segmentation[2]. This approach consists of extracting features from images and training a segmentation model, with numerous recent studies being conducted. Notably, among machine learning-based methods, CNN have gained significant attention for their effective feature extraction capabilities in image segmentation.

One of the representative models in CNN-based image segmentation is the Fully Convolutional Network(FCN). FCN is capable of delivering segmentation results for every pixel within an image. Recently, various studies have been conducted to address the limitations of FCN, such as reduced resolution and slow processing speeds. Notable improvements include models like U-Net[3], DeepLabv3[4], and DINOv2[5]. U-Net builds on the architecture of FCN but enhances accuracy in the segmentation process by symmetrically arranging the encoder and decoder. It incorporates skip connections between the encoder and decoder layers to preserve high-resolution information. DeepLabv3 utilizes atrous convolutions to achieve more precise object boundaries and segmentation. It effectively handles objects of varying sizes within the network and incorporates a module in the final segmentation phase that leverages image-level features to achieve refined object boundaries. DINOv2, a model based on self-supervised learning, employs contrastive learning. It takes advantage of a broader dataset compared to traditional FCN and can flexibly respond to variations in images. The system proposed in this paper performs image segmentation using the DINOv2 model, which has been pre-trained on approximately 142 million images, enabling effective learning of features from a large-scale shipping label image dataset.

OCR is a technology used to identify text within images or videos, consisting of two main processes: text detection and text recognition. Text detection involves extracting regions containing text from images or videos, serving as a preprocessing step necessary for text recognition. Text recognition identifies and converts text from these detected regions into digital text.

Among deep learning-based OCR technologies, methods based on CNN have gained prominence for effectively extracting features from images and videos. Notable models for CNN-based text detection include Efficient and Accurate Scene Text detector(EAST) and Character Region Awareness for Text Detection(CRAFT)[6]. Particularly, CRAFT utilizes both CNN and Recurrent Neural Network(RNN) models to enhance the accuracy of text detection. It extracts features using CNN and processes each pixel sequentially using RNN, specialized for sequential data, to accurately detect the boundaries of text regions.

In terms of text recognition, the Convolutional Recurrent Neural Network(CRNN)[7] represents a leading technology that, like CRAFT, combines CNN with RNN to improve the accuracy of text recognition. Recently, several studies have been conducted to improve the performance of CRNN, with notable advancements including SENet and TrOCR[8]. TrOCR, in particular, has improved upon the CRNN framework by using a Transformer-based network architecture instead of an RNN. This Transformer architecture processes the information all at once, considering the entire sequence, which enables efficient parallelized learning and inference. The CNN extracts feature from images or videos, and the Transformer module, taking these features as input, recognizes text. During this process, the Transformer uses an Attention module to effectively transfer information between CNN and the Transformer, and a Multi-Head Attention module to adapt the system to a variety of texts. This paper employs the TrOCR model in the OCR process.

Named Entity Recognition(NER) is a crucial area of study in the field of natural language processing, involving the identification of named entities within sentences[9]-[11]. For example, the task entails recognizing the category to which a word representing a name belongs. In the sentence "Eugene entered Korea University in 2024," the entities would be recognized as Eugene(name), 2024(time), and Korea University(organization).

Therefore, NER is utilized in various applications, including document summarization, text mining, and analysis. Recently, deep learning-based NER technologies have gained attention. These technologies enable flexible and accurate named entity recognition through automatic feature extraction and extensive data training[12].

A prominent method in deep learning-based NER is the Transformer-based approach. Transformers specialize in processing sequence data, dividing sentences into tokens, and classifying whether each token represents an entity. Transformer-based models like Bidirectional Encoder Representations from Transformers(BERT)[13][14][15] and Efficiently Learning an Encoder that Classifies Token Replacements Accurately(ELECTRA)[16] are effective for named entity recognition because they consider the word information in sentences simultaneously. BERT, a pre-training model for language modeling, employs Masked Language Modeling(MLM) and Next Sentence Prediction(NSP) during its training phase. MLM involves masking random tokens in a sentence and restoring them, while NSP predicts the connectivity between two sentences. This approach demonstrates excellent performance in various natural language processing tasks through transfer learning, although it has the drawback of low training efficiency.

ELECTRA introduces the concept of Generative Adversarial Network(GAN) to natural language processing as a pre-training model. It utilizes a structure where a generator and discriminator networks compete and learn from each other, making the pre-training process efficient. Compared to BERT, ELECTRA operates with a smaller model size and a relatively simpler model structure, leading to faster inference speeds and efficiency.

KoELECTRA[17] is a natural language processing model specialized for the Korean language, based on ELECTRA. Like ELECTRA, it uses a GAN-based learning method, allowing for quick inference with a simple model structure and pre-training adapted to the characteristics of Korean text, effectively handling the context and structure of Korean sentences. In this paper, KoELECTRA is utilized during the named entity recognition stage for masking written Korean on shipping labels.

To collect datasets for model training, we performed web crawling for shipping label images using search engines such as Naver and Google. We specifically collected labels from seven major courier companies in South Korea, which dominate the parcel delivery market, as well as from convenience store courier services. These companies, in order of preference, include Korea Post, CJ Logistics, Coupang, Hanjin Express, Market Kurly, Lotte Logistics, Logen Logistics, and convenience store-based courier services. The collected image dataset was categorized into three types: intact shipping labels, damaged labels that were torn or obscured, and labels with partial masking. Ultimately, the number of shipping label images used for training, arranged in order of company preference, was 255, 194, 193, 96, 68, 56, 36, and 230, respectively.

Prior to training, we generated labeled images for use in image segmentation using the open data annotation platform, Computer Vision Annotation Tool(CVAT)[18]. Intact shipping label images were classified as Type 1, damaged labels as Type 2, and partially masked labels as Type 3. Figure 2 illustrates the proportion of the area occupied by the shipping labels in the dataset, with shipping labels appearing on average in 39% of the images across all types.

Fig. 2. 
Rate of shipping label area distribution

Next, to improve the accuracy of image segmentation training, we augmented the collected image dataset and its boxed version by setting hyperparameters as shown in Table 1. We employed various augmentation techniques including horizontal and vertical flips, rotations, and scaling and cropping methods. As a result, as illustrated in Figure 3, when the original number of images was augmented by three to four times, the highest performance was observed, with a validation dice score of 0.81 after 3000 steps.

Fig. 3. 
Graph of validation dice

Following the preprocessing outcomes in Section 3.1, the parcel shipping label objects within the images were detected using the DINOv2 image segmentation model. We employed a segmentation model rather than an object detection model to precisely determine the boundaries and area ratios of the parcel shipping labels.

Segmentation models, compared to object detection models, can detect the boundaries of objects more finely at the pixel level by dividing them into multiple areas, allowing for accurate calculation of area ratios. The detected area ratios of the parcel shipping label objects are subsequently used as criteria for image classification in the processing stages.

In this study, we considered three segmentation models: YOLOv8[19][20], U-Net, and DINOv2. YOLOv8 is suitable for real-time video segmentation, where processing speed is crucial. Experimental results showed that DINOv2 provided more precise detection of parcel shipping label areas and boundaries compared to U-Net. Therefore, DINOv2 was ultimately utilized.

DINOv2 is a semantic segmentation model that assigns all objects of the same class to the same area. However, this model faced a challenge when multiple shipping labels appeared in an image, as it segmented them into the same area without distinction. To address this, we utilized OpenCV's Contour detection to identify disconnected pixels in the segmentation results, treating them as independent shipping labels. The outcomes of the shipping label area detection in the images are presented in Figure 4.

Fig. 4. 
Shipping label area detection using DINOv2

Upsampling is the process of increasing the sampling frequency of an image or data, thereby expanding or augmenting its size. In particular, upsampling techniques within images can enhance the resolution and enable better discernment of fine details. Specifically, super-resolution technology transforms low-resolution images into high-resolution ones without compromising the details of the original image. In this paper, we utilize the Enhanced Super Resolution Generative Adversarial Networks (Real-ESRGAN)[21], a prominent deep learning-based super-resolution method that employs GAN model.

Real-ESRGAN mimics the real-world photo degradation process more closely through high-order degradation techniques. This involves repeatedly applying the degradation process D, as described in Equation (1), where each iteration of the process D modifies only the hyperparameters of the same degradation sequence. Additionally, the degradation process, as shown in Equation (2), sequentially applies blur, resize, noise, and JPEG compression to the ground-truth input image y.

In the proposed system, Real-ESRGAN is used to enhance the resolution of the shipping label images, thereby facilitating smoother text detection and extraction of positional coordinates. The process is as follows: Initially, the minimum and maximum x and y coordinates of the shipping label areas detected by DINOv2 are used to define a rectangular region enclosing the label. Subsequently, this rectangular area is extracted from the entire image, and the cropped shipping label image is input into the Real-ESRGAN model to generate a high-resolution output. Figure 5 compares the original shipping label images with those improved using super-resolution technology.

Fig. 5. 
Super-resolution x2 using real-ESRGAN

Prior to text extraction from the high-quality cropped images of shipping labels obtained through Real-ESRGAN, we used the CRAFT model to detect the locations of text. CRAFT enhances the accuracy of character detection by combining CNN and RNN technologies, where the CNN extracts feature from the image and the RNN identifies the boundaries of character regions.

The application process of the CRAFT algorithm is as follows: Given a specific word sample 'w' from the training data, the area of the word is denoted as 'R(w)', and the length of the word as 'l(w)'. By segmenting the word into characters, the boxes for each character and their total length 'lc(w)' are estimated. Subsequently, a confidence score for each word sample 'w' is calculated using the following equation (3).

In the given image, a confidence score 'sconf(w)' is assigned to each pixel 'f' within the word region 'R(w)', and different values are assigned to all other pixels to generate a confidence map 'Sc’. The overall objective 'L' is to compute the sum of the values of the confidence map 'Sc(p)' for each pixel 'p', and the product of the sum of the squared differences between this and the ideal area score 'Sr(p)' and affinity score ‘Sa(p)’.

In equation (4), 'Sr*p' and 'Sa*p' represent the ideal area and affinity scores, respectively, while 'Sr(p)' and 'Sa(p)' denote the model's predictions for these scores. When training with synthetic data, since the actual ground truth is available, the confidence map 'Sc(p)' for every pixel 'p' is set to 1. The coordinates of text regions obtained through the CRAFT algorithm are then used in the text recognition and masking stages. The heatmap showing the regions, locations, and boundaries of characters detected by CRAFT, along with the extracted text areas, is presented in Figure 6.

Fig. 6. 
Text recognition using CRAFT

In this study, we considered two text recognition models: CRNN and TrOCR. Both models were trained using 9,000 images of printed Korean text and 1,000 images of handwritten Korean text provided by AI-Hub[22]. TrOCR demonstrated superior performance; therefore, it was selected for use in this research. Figure 7(a) illustrates text locations obtained using CRAFT, which TrOCR utilizes to identify text positions and extract each text as shown in Figure 7(b). While cropped images generated by Real-ESRGAN showed excellent performance in confirming text locations, the quality alterations in these cropped images led to decreased character recognition rates when used for text recognition. Therefore, original images were used for text recognition purposes[23].

Fig. 7. 
Text Extraction using CRAFT and TrOCR

TrOCR consists of an encoder that converts images into vectors and a decoder that translates these vectors into words. It takes an image containing text as input (3*H0*W0), resizes it, and then splits it into patches of size p*p. Next, as described in Equation (5), each patch is embedded and then linearly projected to the Transformer's hidden size D. The D vectors are fed into the decoder, where the word corresponding to each vector h_i is identified. Equation (6) represents the probability that the i-th vector's word is the j-th word among V possible words.

To determine whether the text information obtained from the TrOCR text recognition model contains personal data such as names, addresses, and phone numbers, we applied the fine-tuned named entity recognition model, KoELECTRA-small-v3-modu-ner. This model was pretrained on a Korean dataset categorized into 15 tag sets according to the broad classification standards of the Korea Information and Communication Technology Association(TTA). It takes text information as input and assigns classes to words(named entities) corresponding to people, roads, buildings, etc., according to predefined classification standards, as shown in Figure 8(a). In our system, the named entity recognition model classifies personal information in the text of shipping labels—such as names, addresses, and phone numbers—as Class 1, while all other text is classified as Class 0. For names and addresses, the model uses tags for people(PS), locations(LC), and areas(AR) from its broad classification criteria to classify them as Class 1. Additionally, numerical information like phone numbers and shipping label numbers is classified as Class 1 through a separate definition for masking purposes. Figure 8(b) illustrates the results of the text recognition model TrOCR and its subsequent classification using the KoELECTRA named entity recognition model. Using the class information obtained from this process along with the text locations previously determined by CRAFT, masking is performed on the original images.

Fig. 8. 
Classification of extracted texts using KoELECTRA

Image segmentation takes the images processed during data collection and preprocessing as input. The models considered for the image segmentation include YOLOv8, U-Net, and DINOv2. To compare the performance of these models, we compared the original images(ground truth) with the mask images predicted by the three models.

The hyperparameters used were as follows: all three models employed the SGD optimizer; YOLOv8 used a default learning rate of 0.01, while U-Net and DINOv2 used 4e-5. The batch sizes were set at 32, 16, and 8 for YOLOv8, U-Net, and DINOv2, respectively. YOLOv8 utilized a model pretrained on the COCO 2017(Common Objects in Context 2017) dataset[24], U-Net was used as a basic model without pre-training, and DINOv2 was simulated with a pretrained backbone through fine-tuning.

Performance comparison among the three models was conducted using metrics such as dice score and mAP(mean Average Precision), as outlined in Table 2. The dice score, which measures the similarity between the mask generated by the model and the actual image mask, is calculated by dividing the size of the overlapping area by the total area size. Scores closer to 1 indicate superior model performance. The dice scores for YOLOv8, U-Net, and DINOv2 were 0.74, 0.68, and 0.82, respectively, indicating that DINOv2 performed the best. mAP, which calculates the average over all classes by considering the area under the curve of precision and recall for each class, also indicates better object detection performance as it approaches 1. For a more accurate performance comparison, mAP was measured incrementally from an IoU of 0.5 to 0.95 by 0.05, alongside mAP 50, which assumes detection only if IoU is above 0.5. The IoU indicates the degree of overlap between the predicted bounding box and the ground truth bounding box. As shown in Table 2, the mAP scores for YOLOv8, U-Net, and DINOv2 were 0.68, 0.64, and 0.74, respectively, and the mAP 50 scores were 0.81, 0.77, and 0.86, respectively, with DINOv2 showing the highest performance in both metrics. Additionally, as depicted in Figure 9, DINOv2 was able to extract the borders of the shipping labels most clearly in the images. Therefore, DINOv2, having demonstrated the highest performance across all metrics, was used for image segmentation in this system.

Fig. 9. 
Comparison of image segmentation between U-Net and DINOv2

For the character recognition model, simulations were conducted using both the CRNN and TrOCR models on images resulting from the text area detection model. To compare the performance of the two models, we analyzed three images(ground truth) as shown in Figure 10, and compared the results applied by each model. The hyperparameters utilized in the experiment were as follows: both CRNN and TrOCR utilized the Adam optimizer with a learning rate set at 1e-3. Additionally, the batch sizes were set at 80 for CRNN and 20 for TrOCR, with only the TrOCR model undergoing fine-tuning. Experimental results, as shown in Table 3, indicated that TrOCR exhibited superior character recognition accuracy across the three images compared to CRNN. Consequently, the TrOCR model was selected for text recognition.

Fig. 10. 
Input images of text recognition model

To determine whether the text extracted by TrOCR contains personal information, named entity recognition was performed using the BERT and KoELECTRA models through binary classification. Both models were trained using Binary Cross Entropy Loss(BCE Loss). Upon measuring the validation accuracy, as shown in Table 4, BERT achieved an accuracy of 0.68, while KoELECTRA achieved a higher accuracy of 0.82.

To measure the performance of the final model, we conducted simulations integrating the models introduced from Sections 4.1 to 4.3 and verified their accuracy. Simulations were performed on six shipping label images containing personal information. During the image segmentation, image quality enhancement, and text detection stages, the models DINOv2, Real-ESRGAN, and CRAFT were applied based on experimental and model characteristics. For text recognition and named entity recognition, it was crucial for our model to accurately recognize the preprocessed images and texts and identify personal information. Therefore, we combined two text recognition models and two named entity recognition models, conducting a total of four simulations. Accuracy was measured based on the ratio of personal information correctly identified from the text images, as extracted by the CRAFT model. As shown in Table 5, the combination of TrOCR and KoELECTRA as the text recognition and named entity recognition models, respectively, achieved the highest accuracy of 0.8124 among the four scenarios.

Personal information masking, or de-identification, involves techniques such as mosaicking and blurring to obscure personal text data. In this study, instead of distorting images with a fixed-size pixel mosaic, blurring was applied to preserve the natural appearance of the images. The Gaussian Blur used in our system averages the pixel values around each pixel in the input image by assigning weights to the surrounding pixels, thus effectively masking the data. Figsure 11 demonstrates the experimental results of identifying the positions of personal information text within shipping labels and applying masking based on these locations.

Fig. 11. 
Masking stage

The proposed system offers more functionalities compared to existing systems from a technical perspective, as shown in Table 6. First, it provides a clear classification of the recognized text according to the type of personal information (e.g., name, phone number, address). Second, the system applies the CRAFT technique to accurately detect the position of unstructured, variable data and adapt to different text sizes. Third, to enable precise correction and detection from images distorted at various angles, the system incorporates Real-ESRGAN generation technology and semantic segmentation. Overall, the system demonstrated a 19.9% improvement in recognition rate for distorted images compared to Tesseract OCR, with a successful classification and masking rate of 81.2% for various types of personal information.