foreword

This is a part of the data set that I have used in my work and study. There are several categories such as semantic segmentation, target recognition, and portrait matting. This is just a part of the data set I have used. There are a small part of these data sets. The source is from the Internet, and a large part of it is collected by myself.

1. Semantic Segmentation

1. Book edge segmentation

This data set marks the middle line and edge data set of the book. The data is marked with labelme. The data set has 2500 images, which are used to train the edge detection and center line recognition of document scanning. I have trained with ENet before, and the effect is quite good. An example of dataset labeling is as follows:

2. Human face skin

2.1 Wrinkles and eye bags (164 images)

The data set marks the location of wrinkles and bags under the eyes, and the data set has a converted mask image, which can be used for skin wrinkle detection and eye bag detection and segmentation in medical beauty. For training methods and effects, please refer to my previous blog:

Realization of Wrinkle Detection, Positioning and Segmentation of Face Image Based on Semantic Segmentation

2.2 Acne, erythema, freckles (221 photos)

The dataset contains labels such as acne, erythema, and freckles, which can be used to train real instance segmentation, and each target has a unique ID.

acne

erythema

freckle

2.3 Pores

The pore data set contains only pores and can only be used for training semantic segmentation.

3. The photo file is clear and binary

The data set is used for clear binarization of document photos. In some apps, this function is called ink-saving mode. The samples are used to train the filter. The training effect can be used in my previous blog:

Use deep learning to solve the problem of binarization of complex backgrounds in photographed documents_How to binarize images with darker backgrounds

4. Shadow detection and segmentation of photographed documents

The data set is marked with labelme, marking the shadow part of the photographed document, which is used for shadow detection and removal of photographed document scanning. There are a total of 1000 images.

5. Test sheet

The test sheet is collected by taking pictures, and the data volume is more than 2W, including handwritten and machine-printed. The data object only marks the edge and the form line on the test, and the form line is used for form extraction. The annotation tool is labelme.

About 30,000 handwritten test sheets have also been collected, which can be used for doctor's handwritten OCR recognition:

2. Target recognition

1.1 mobile phone

The marked is the mobile phone, the data format is xml and the txt format of the converted yolo, and there are 2000 pieces.

1.2 Hard hat

The data is marked with heads wearing safety helmets and heads without safety helmets. The data format is xml and converted yolo txt format. There are a total of 7,800 images. The training effect can refer to my previous blog:

Hard hat wearing detection - from data processing, training data to model deployment (with data set, training code, C++ model deployment code that can use GPU)_c++ ipc to detect hard hat

1.3 Fireworks

The data is marked with two tags of open flame and smoke. The data format is xml and converted yolo txt format. There are a total of 7600 images. The training effect can refer to my previous blog:

Fireworks detection based on Yolov5 - model training and C++ deployment - yolov5 c++ deployment

1.4 Smoking

The data is marked with two tags of head and smoking. The data format is xml and converted yolo txt format. There are a total of 4800 images.

1.5 Photographic documents

Document object detection detects two objects, a double-open book and a single document. The labeling tool is labelImg, and the data format is xml and yolo's txt.

single document

double book

1.6 ID card

The data set includes the front and back of the ID card. The front of the ID card is marked with the front and head, and the back of the ID card is marked with the back and the national emblem. There are four targets in total. The labeling tool is labelImg, and the data format is xml and yolo's txt. The training effect can refer to my subsequent blog:

Document photo scanning - based on C++ and deep neural network to realize document recognition and scanning and restore documents to A4 paper 1 to 1

1.7 Bank Card

The data set includes the front and back of the bank, UnionPay logo, IC chip, a total of four targets. The labeling tool is labelImg, and the data format is xml and yolo's txt.

1.8 Passport

The passport is marked with the first page of the passport and the portrait of the person, a total of two goals. The labeling tool is labelImg, and the data format is xml and yolo's txt.

1.9 Hong Kong and Macau Pass

The Hong Kong and Macau Pass is marked with three targets: the front, the profile picture, and the back. The labeling tool is labelImg, and the data format is xml and yolo's txt.

1.10 Residence Permit

The residence permit marked the front face and head portrait, a total of two targets. The labeling tool is labelImg, and the data format is xml and yolo's txt.

1.11 Driver's license

The driver's license is marked with the front face and head portrait, a total of two targets. The labeling tool is labelImg, and the data format is xml and yolo's txt.

1.12 Social Security Card

The social security card is marked with the profile picture, the UnionPay logo, the IC card, the front of the social security card, the back of the social security card, and the national emblem. The labeling tool is labelImg, and the data format is xml and yolo's txt.

3. Cutout

1. Portrait Cutout

There are half-length cutouts for portrait cutouts, about 50,000 pieces, and the format is a mask image, which was previously used for smart ID photos.

To be updated......................

Datasets - personally collected and used datasets