AN INTELLIGENT SYSTEM FOR DOG BREED RECOGNITION USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Introduction. Convolutional neural networks (CNNs) have established themselves
as a powerful tool for image analysis. Inspired by biological visual processing, they automatically
detect complex patterns in data by sequentially extracting increasingly abstract features. The growing
demand for AI-powered interactive applications makes automated image recognition systems
particularly relevant across a wide range of domains, including veterinary practice, animal shelters,
and mobile applications.
The Purpose of the article is to theoretically justify and practically demonstrate the use of
convolutional neural networks for dog breed identification, and to describe the full development cycle
of a CNN-based web application – from dataset preparation to deployment.
Results. The article presents a review of image classification methods based on deep learning,
with a focus on the architecture and training principles of CNNs. The roles of convolutional layers,
pooling layers, the ReLU activation function, and the backpropagation algorithm are discussed. Three
Python libraries central to the project are reviewed: Pandas for data preprocessing, TensorFlow for
building and training the neural network model, and Streamlit for rapid development and deployment
of the web interface.
The practical implementation covers all stages of system development. A dataset of 1,888
images across 21 dog breeds was collected using the bing-image-downloader library and manually
filtered. The images were split into training (90%) and validation (10%) sets. The CNN architecture
follows a sequential structure consisting of four convolutional-pooling blocks with 16, 32, 64, and 128
filters respectively, followed by two fully connected layers (1,024 and 256 neurons) with Dropout
regularization (p = 0.2), and a Softmax output layer with 21 neurons. The model was trained for 30
epochs and achieved a classification accuracy of 93.64% on the test set. Practical testing confirmed
high accuracy for breeds present in the training data (e.g., Husky – 94.08%, Jack Russell Terrier –
99.98%) and predictable probabilistic distribution for breeds outside the dataset.
The web application was deployed via Streamlit Cloud integrated with a GitHub repository,
providing a user-friendly interface for uploading images and displaying the top five predicted breeds
with corresponding probabilities.
Conclusion. Convolutional neural networks are an effective tool for dog breed classification
tasks. The combination of TensorFlow for model development and Streamlit for interface deployment
provides an efficient and accessible pipeline for building machine learning–based web applications.
The proposed approach can be readily extended to other image recognition tasks and illustrates the
practical potential of deep learning in real-world automated recognition systems.