ISSN (Online) 2581-9429

IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 6, May 2022

Impact Factor: 6.252

Automated HTML Code Generation from Hand

Drawn Images using Machine Learning Methods
Gayatri Vitkare1, Rutuja Jejurkar2, Sammyaka Kamble3, Yogeshwari Thakare4, Prof. P. A Lahare5
Student, Department of Information Technology1,2,3,4
Professor, Department of Information Technology5
Pune Vidyarthi Griha's College of Engineering & S. S. Dhamankar Institute of Management, Nashik, India

Abstract: The production of individual web page mock-ups, which can be done by hand or with the aid
of graphic design and professional mock-up creation tools, is the first stage in the website design process.
The mock-ups are then converted into structured HTML or similar mark-up code by software engineers.
Typically, this method is performed multiple times until the required template is achieved. The purpose of
this review is to make the process of developing code from hand-drawn mock-ups more automated. Hand-
drawn mock-ups are processed using computer vision techniques, and the recommended system is built
using deep learning techniques. The building of a preliminary drawing of each web page, which can be
done using design tools or by hand. After that, corresponding code for the web page draught is written.
This procedure is difficult, expensive, and time-consuming. Consequently, the suggested system will
automate this operation. A hand-drawn drawing of a form is provided as input, which is analysed, and
several components revealed. After the components have been discovered, deep learning CNN techniques
are used to crop and recognise them. When the matching component is identified.

Keywords: Machine learning with convolution neural networks (CNN), Object detection, object
recognition, HTML Code generation, HTML (Hypertext markup language)

I. INTRODUCTION
Because of today's technological advancements, the relevance of Internet web pages has expanded significantly.
Nowadays, websites represent the personalities of nations, institutions, communities, and individuals. There are websites
for practically any subject, from information to social work, games to training, and so on. Companies' websites are
brought to the forefront for financial objectives, such as product promotion or advertising. Developers are in charge of
creating client-side software based on a designer's mock-up of the Graphical User Interface (GUI). The implementation
of GUI code, on the other hand, requires time and diverts engineers' attention away from the product's core functionality
and logic. In addition, the computer Furthermore, each target runtime system's computer languages for designing such
GUIs are separate, resulting in onerous and repetitive effort when the product is meant to function on several platforms
using native technologies. In this research, we present a model that has been trained end-to-end with stochastic capable
to reduce to construct variable-length tokens strings from a single GUI picture given input, yet staying technically and
informational. An algorithm was created in this work to automatically produce HTML code for a hand-drawn mock-up
of a web page. Its goal is to detect the mock-up picture's components and encode them according the web page structure.

II. LITERATURE SURVEY

A large gap occurs between graphic designers' conceptual designs and functional UI code while building the UI code of
a mobile application. Programmers often close this gap by rewriting the conceptual drawings in code, which is time-
consuming and costly. To fill this need, we pioneered the first method for automatically reverse engineering mobile app
user interfaces (REMAUI). REMAUI uses computer vision and OCR methods to identify UI components on a given
input image. REMAUI-generated UIs were close to the originals in our tests on 488 screenshots of over 100 major third-
party programmes. We intend to expand the export process to more platforms, including iOS and cross-platform
JavaScript frameworks. Currently, REMAUI separates each input screen into its own programme. We want to provide a
graphical notation that users can use to connect many inputs screen designs, which REMAUI can then utilise to produce
a single application with several screens and matching screen transitions. REMAUI will be integrated with tools that
Copyright to IJARSCT DOI: 10.48175/IJARSCT-4222 146
www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 6, May 2022

Impact Factor: 6.252
produce mobile app functionality using keyword-based code search [44] or high-level models.) We want to index a
screenshot corpus by executing REMAUI on it and saving the intermediate results of REMAUI. By exposing this index
through a query interface, a user might look for screenshots based on their structure and features.
Building classifiers for picture detection and recognition on multiple datasets using various machine learning had
advanced significantly in recent years. Deep learning, in particular, has improved accuracy across a variety of datasets.
The below are summaries of some of the works. SVM (support vector machines), KNN (K-nearest Neighbour), and
CNN were used to evaluate the performance on MNIST datasets (convolutional neural networks). On that platform, the
Multilayer Perceptron didn't function well since it didn't attain the global minimum but rather stayed caught in the local
optimum and couldn't reliably distinguish digits 9 and 6.Other classifiers performed well, and it was decided that by
building the model on the Keras platform, performance on CNN may be enhanced. On the MNIST dataset, implement
DNN (deep neural networks), DBF (deep belief networks), and CNN (convolutional neural networks) and compare the
results. According to the research, DNN fared the best, with an accuracy of 98.08 percent, while others had error rates
and execution times that differed.
The study was classified into four categories. In the first stage, image processing methods such as erosion, dilation, and
contour detection were being used to search for things in the input image. Following that, the identified objects were
cropped, and the associated components were labelled using the trained CNN model. Finally, the model's output was
transformed to HTML code using the HTML Builder script. Once the input data has already been read, it is transformed
to grey scale style. Gaussian Blur was then reapplied to them using a 3x3 rectangle kernel. Following the thresholding
step, the contour detection technique was used to create a rectangle in order to identify the objects basis of morphological
changes. In this approach, the components of the uploaded picture have been detected. Cropping was used to transfer the
identified components to the CNN model. In the phases of morphological features, 8 iteration dilation with a 4x4
rectangle kernel was implemented. With a 3x3 ellipse kernel, the erosion process was used for 4 repetitions, and the
dilation process was used for 2 iterations with a 1x10 rectangle shaped kernel. Finally, a 10x1 rectangle shaped kernel
was used in the dilatation operation. The model was trained using our component dataset chunks. It is made up of four
major types of components, as previously stated: textboxes, drop-down menus, buttons, and checkboxes. After the model
was trained, the loss function was trained for 200 epochs using Binary Cross Entropy and RMS Prop methodologies with
a batch size of 64. After that, the component recognition procedure used the cropped elements from the preceding stage
as input. For feature extraction, we utilized several convolution layers with 4x4 kernels, similar to our CNN Model, and
thereafter max pooling operations using 2x2 kernels. We apply the BiLSTM layer to catch the relationship of the
collected features after the feature vectorization procedure. After all, we utilised a 20% ratio of Full Connected Layers
to Dropout Layers to achieve the classification aim. Recognized components were successfully translated into HTML
code via the bootstrap framework. It was performed with the help of the coordinates from the result of the contour finding
algorithms.

III. PROPOSED METHODOLOGY

The technology is used to automate the web page creation process. The system is given a hand-drawn image of a web
form as input, and corresponding HTML code is to be created.There are four key phases to any process of developing
robotic HTML code. The first step is object detection, which involves scanning the picture and using image processing
algorithms such as LSTM, DSL, and so on. After then, the objects which have been identified are snipped. Eventually,
the outcome of the model is converted to HTML code

3.1 Pix2Code
The method evaluates a snapshot of a graphical user interface (i.e., the layout of a web application interface) and
determines what the underlying code looks like through an iterative process. Making computer code from a GUI
screenshot is analogous to writing English written explanations from scene images. In both situations, we wish to
generate variable-length token strings from pixel values. As a result, our problem may be divided into three pieces. To
begin, understanding a given scene (in this example, the GUI display) and inferring the things there is a computer vision
challenge, their IDs, two positions, and poses (i.e., buttons, labels, element containers). Secondly, reading and writing
syntax correct language (in this case, computer code). It's a difficulty to represent language using conceptually correct
examples. Lastly, and used the latent variables identified via scene translation, combine the responses including both
Copyright to IJARSCT DOI: 10.48175/IJARSCT-4222 147
www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 6, May 2022

Impact Factor: 6.252
sub-problems and provide corresponding textual descriptions (i.e., computer code rather than English) of the things
represented by all these variables. The following parts make up Pix2code:
 LSTM (Long-Short Term Memory)
 CNN (Convolutional Neural Network)
 DSL (Domain specific Language)

Figure: Proposed System Architecture

Detection of Objects
Textboxes, buttons, and labels are among the components of Dataset [5.] Contour detection and morphological
transformation are used to detect these elements in a picture. Outside the component, contour detection is employed to
draw boundaries. Using morphological erosion, little things are eliminated, leaving only substantial objects. Objects are
made more apparent by filling gaps in them using morphological dilatation. These strategies improve object visibility
and clarity.

Object Recognition
Convolutional Neural Network has used to train the dataset. Input image of web form then is given to this training set
which recognises components inside an image. The image from the online form is then sent to this trained model, which
recognises image components. Object recognition is a challenging computer vision problem that necessitates estimating
both of the position and kind of entities seen in a picture.

Training data set

A dataset is a grouping of images with textboxes, buttons, and labels. A dataset then used train a machine learning
technique is referred to as a training model.The training model is used to run the algorithm on the input data and compare
the processed result to the predicted outcome.The model was derived based upon the results with this connection.

CNN
The elements were labelled using the trained CNN model after the recognised objects were clipped. Before even being
fed through into CNN, the input pictures are scaled to 256*256 pixels (the aspect ratio is not kept) and the pixel values
are balanced. There will be no further processing. We simply utilised 3*3 receptive fields convolved with stride 1 to
encode each input image to a fixed-size output vector. To descending utilizing max-pooling, these processes are done
again.

HTML Code Generation

For the detected components, HTML code is generated using the. Gui file created in the previous phase. The goal is to
encode the web page's recognised components. First, header and footer templates are generated. Lastly the system creates
web page file after generating the HTML code.

Copyright to IJARSCT DOI: 10.48175/IJARSCT-4222 148

www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 6, May 2022

Impact Factor: 6.252
IV. RESULTS
The goal of this research has been accomplished. We were able to generate HTML code from such a mock-up image
with the addition of color buttons. All of the approaches are the foundational factors that have led to the current and
planned system.

Fig. Input Image Fig.HTML Code

Fig. Output
V. FUTURE WORK
 Expanding system adaptability by allowing users to generate front-end designs with a range of image types.
 Improving the system so that more appealing designs can be created.
 Adding a new function that allows users to customise the CSS for the website's front end to improve system
quality.

VI. CONCLUSION
A critical aspect has been converting website mock-ups into mark-up code with less time and development expense. We
built a technique in this work that receives web page mock-ups, processes them, and generates structured HTML code.
A collection of images was used, which included several mock-ups of web page architectures. Future research and
expertise are needed to improve the code generation's efficiency. Such apps are needed in the industry and in our daily
lives since every firm insists on making their laborious task easier and more efficient. The above study presents a system
that turns picture data into an HTML webpage automatically.

Copyright to IJARSCT DOI: 10.48175/IJARSCT-4222 149

www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 2, Issue 6, May 2022

Impact Factor: 6.252
REFERENCES
[1]. Automatic HTML Code Generation from Mock-up Images Using Machine Learning Techniques978-1-7281-
1013-4/19/$31.00 © 2019 IEEE
[2]. Convolutional Neural Network (CNN) for Image Detection and Recognition 978-1-5386-6373-8/18/$31.00
©2018 IEEE
[3]. Reverse Engineering Mobile Application User Interfaces with REMAUI 978-1-5090-0025-8/15 $31.00 © 2015
IEEE DOI 10.1109/ASE.2015.32
[4]. pix2code: Generating Code from a Graphical User Interface Screenshot arXiv:1705.07962v2 [cs. LG] 19 Sep
2017
[5]. Norhidayu binti Abdul Hamid, Nilam Nur Binti Amir Sjarif, “Handwritten Recognition Using SVM, KNN and
Neural Network”, www.arxiv.org/ftp/arxiv/papers/1702/1702.00723
[6]. Mahmoud M. Abu Ghosh; Ashraf Y. Maghari, “A Comparative Study on Handwriting Digit Recognition Using
Neural Networks”, IEEE, 2017
[7]. T. A. Nguyen and C. Csallner, “Reverse Engineering Mobile Application User Interfaces with REMAUI (T),”
in 2015 30th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, nov
2015, pp. 248–259. [Online]. Available: http: //ieeexplore.ieee.org/document/7372013/
[8]. S. Natarajan and C. Csallner, “P2A: A Tool for Converting Pixels to Animated Mobile Application User
Interfaces,” Proceedings of the 5th International Conference on Mobile Software Engineering and Systems -
MOBILESoft ’18, pp. 224–235, 2018. [Online]. Available: http://dl.acm.org/citation.cfm?doid=31
97231.3197249.

BIOGRAPHY
1. Gayatri Vitkare,Under Graduate Student,PVGCOE & SSDIOM,Nashik, Maharashtra, India
2. Rutuja Jejurkar,Under Graduate Student,PVGCOE & SSDIOM,Nashik, Maharashtra, India
3. Yogeshwari Thakare,Under Graduate Student,PVGCOE & SSDIOM,Nashik, Maharashtra, India
4. Sammyaka Kamble,Under Graduate Student,PVGCOE & SSDIOM,Nashik, Maharashtra, India
5. P.A.Lahare, Professor,PVGCOE & SSDIOM,Nashik, Maharashtra, India

Copyright to IJARSCT DOI: 10.48175/IJARSCT-4222 150

www.ijarsct.co.in