کنفرانس مهندسی برق ایران

صفحه اصلی / سی و دومین کنفرانس بین المللی مهندسی برق

A Deep Learning-Based Model for House Number Detection And Recognition

نویسندگان :

Roghaiyeh Tayefeh Younesi¹ Jafar Tanha² Samaneh Namvar³ Sahar Hassanzadeh Mostafaei⁴

1- دانشگاه تبریز 2- دانشگاه تبریز 3- دانشگاه تبریز 4- دانشگاه تبریز

کلمات کلیدی :

Image Recognition،Image Detection،Data Augmentation،CNN،LSTM،Multi Digit

چکیده :

Abstract— Detection and recognition of information from natural images pose significant challenges in computer vision, with far-reaching implications for future applications. In recent years, the application of deep learning techniques to real-world image datasets has yielded notable achievements in the realms of recognition, detection, and pattern recognition. In this paper, we specifically tackle the challenge of number detection and recognition in real-world scenes by proposing deep learning models on the Street View House Numbers (SVHN) dataset. In the proposed models, to boost accuracy, we applied preprocessing steps to the training dataset. These steps included data augmentation techniques such as resizing, random rotation, random horizontal flip, angle degree changes, and optimization of hyperparameters and model layers. In the initial model, we utilized a fully connected Convolutional Neural Network (CNN) model on sequences of digit images, achieving an impressive accuracy of 95 percent. Subsequently, a Convolutional-Long Short-Term Memory (CNN-LSTM) model was designed for temporal information modeling, utilizing a combination of CNN and LSTM layers that also achieved an accuracy of 93 percent. These models demonstrate high performance in recognizing numbers in complex and real-world environments. Our results underscore the significant enhancement in the accuracy of number recognition in real-world images achieved on the SVHN dataset by combining CNN models with data augmentation. We also compare the results of our proposed models with other state-of-the-art methods.