Image recognition using machine learning

ANKITA SINGH RATHORE

Department of Computer Science & Engineering
[email protected]

AURANGJEB ALAM
Department of Artificial Intelligence & Data Science

[email protected]

ARYA COLLEGE OF ENGINEERING, JAIPUR, RAJASTHAN, INDIA

Abstract
The essential facet of image processing for machine learning involves image recognition
without any human intervention throughout the process. This research delves into the methodology
of image classification employing an image-based backend. A substantial number of images
portraying both cats and dogs are collected and subsequently partitioned into distinct categories for
the test and training datasets essential for our learning model. The outcomes are derived through the
utilization of a bespoke neural network featuring Convolutional Neural Networks architecture,
facilitated by the Keras API.

Keyword:
Image analysis, automated image processing, machine-driven image recognition, autonomous
classification, dataset segmentation, neural network customization, CNN architecture, Keras
framework, computer vision, animal categorization, image dataset collection, model training,
experimental results.

1.Introduction
Image classification has emerged as a pivotal tool bridging the gap between computer
vision and human perception, achieved through the training of computers with vast datasets.
Artificial Intelligence (AI) has long been a focal point of scientific and engineering endeavors, aimed
at enabling machines to comprehend and navigate our world to serve humanity effectively. Central to
this pursuit is the field of computer vision, which focuses on enabling computers to interpret visual
information such as images and videos. In the early stages of AI research from the 1950s to the
1980s, manual instructions were provided to computers for image recognition, employing traditional
algorithms known as Expert Systems. These systems necessitated human intervention in identifying
and representing features mathematically, resulting in a laborious process due to the multitude of
possible representations for objects and scenes. With the advent of Machine Learning in the 1990s, a
paradigm shift occurred, enabling computers to learn to recognize scenes and objects autonomously,
akin to how a child learns from its environment through exploration. This shift from
instructing to learning has paved the way for computers to discern a wide array of scenes and objects
independently.
Section II provides an overview of the basic artificial neural network, while Section III delves into
Convolutional Neural Networks. The implementation details and resultant findings are expounded
upon in Section IV, followed by conclusions drawn in Section V. Finally, the references are furnished
at the conclusion of the document.

2.Artificial neural network

An artificial neural network comprises interconnected hardware components, often augmented or
segregated by software systems, mirroring the functioning of neurons within the human brain.
Introducing a multi-layered neural network serves as a potential solution to enhance performance.
Effective training of such networks mandates a substantial number of image samples, at least nine
times greater than the parameters necessitated for refining classical classification, ensuring optimal
resolution. The architecture and operations of neural networks are crafted to mimic associative
memory, wherein learning occurs through the processing of example inputs and corresponding
outputs, establishing weighted connections stored within the network's data structure.

In training the model, inputs traverse through hidden layers, undergoing custom grid image
processing to extract pertinent data from distinct sections, subsequently informing the network of its
output. The complexity of neural networks is articulated in terms of the layers involved in input-
output production and the network's depth. Notably, Convolutional Neural Networks (CNNs) have
garnered significant attention for their adeptness in implementing genetic algorithms within hidden
layers, incorporating techniques such as pooling and padding to prepare data from test datasets for
integration into the training model.

Fig.1
3. Convolutional Neural Network
Convolutional Neural Networks (CNNs or ConvNets) represent a pivotal class of deep
learning architectures primarily utilized for analyzing visual data. Renowned for their
shiftinvariant nature and shared-weights structure, they are adept at processing images and
videos, boasting applications across diverse domains such as image classification, medical
imaging, recommender systems, natural language processing, and financial analytics. CNNs
operate by overlaying a 3x3 cell matrix onto input images, transforming them into feature
maps consisting of binary values (1s and 0s). This process iterates across the entire image,
progressively refining feature detection in subsequent layers.

Fig.2

During training, the network discerns crucial features essential for effective image scanning
and categorization, refining its feature detectors accordingly. Often, these features may not
be discernible to human observers, underscoring the remarkable utility of convolutional
neural networks (CNNs). Through extensive training iterations, CNNs can vastly surpass
human capabilities in image processing, making significant strides in accuracy and
efficiency.

4. Implementation, Results and Discussion

In our implementation, we curated a dataset comprising approximately 24,000 images of
cats and dogs, incorporating variations such as rotations and scaling to diversify the training
set. To ensure robust evaluation, we partitioned the dataset into a training set encompassing
90% of the data and a separate testing set. Employing TensorFlow as our backend,
leveraging the computational prowess of a discrete GPU (specifically, the GTX 1050ti with
4GB of memory), we embarked on training our model over 10 epochs. The initial layer
produced an output volume size of 55x55x96, with subsequent layers building upon the
feature maps generated by their predecessors. The model architecture entailed the
application of dense filters with a matrix size of 128x128, yielding an accuracy of 77.8%.
However, recognizing the potential for enhancement, we augmented the model complexity
by increasing the number of layers and neurons, extending the training cycles to 20 while
concurrently reducing the learning rate from 0.001 to 0.0001. This modification, coupled
with a higher-level filter of 256, propelled the accuracy to 88%. Further refinements,
including adjustments to filter size and learning rate, led to a notable increase in accuracy to
93%. Despite achieving promising results on the training set, the true test lay in evaluating
the model's performance on the test dataset. Extending the training cycles further, albeit at
the cost of increased computational resources and time, we observed a remarkable accuracy
of 97.3%. These findings underscore the efficacy of our approach in achieving high
accuracy in image classification tasks, albeit with significant computational demands.

Fig.3

Fig.4

Fig.5

Fig.6

Fig.7

In fig.3 this dog’s images recognise the breeds of dogs. fig.4 this dog’s images recognise the
breeds of cats. Fig.5 this butterfly images recognise the types of butterflies. fig.6 this cows
images recognise the breeds of cow. fig.7 this rabbits images recognise the breeds of rabbit

Our observations revealed some variability in results; however, on average, the accuracy
consistently ranged between 90-95% when employing a layer filter size of 256. This
suggests that leveraging more potent hardware could potentially yield even greater results.
Additionally, expanding the dataset to encompass a wider array of categories beyond just
two classes could further enhance the model's performance. By embracing these strategies,
we anticipate achieving even higher accuracies and bolstering the model's capabilities in
handling more complex classification tasks.
5. Conclusion
In conclusion, our experimentation with random images yielded successful results. We
sourced the image dataset directly from the Google repository and employed a
convolutional neural network in conjunction with Keras for classification tasks. Through
our experiments, we noted that the model effectively classified images even when
subjected to variations such as scaling, trimming, or rotation, generating entirely new
inputs. This underscores the efficacy of deep learning algorithms in handling diverse and
complex image classification tasks with robustness and accuracy.

6. REFERENCES
1. Elsken, Metzen, and Hutter's studies provide insights into efficient architecture
search for convolutional neural networks (CNNs), supporting the methodology
employed in this research.
2. Springenberg et al.'s work on all convolutional net architecture contributes to
understanding the effectiveness of streamlined CNN designs.
3. Romanuke's research on the appropriate allocation of Rectified Linear Units
(ReLUs) in CNNs adds valuable insights into activation functions and network
optimization.
4. Foundational work by Bengio et al. on greedy layer-wise training of deep networks
establishes fundamental principles underpinning deep learning methodologies.
5. Documentation from TensorFlow and Keras libraries serves as essential resources
for implementing and understanding CNN models, validating the methodology and
results presented in the research.
6. The introduction of TensorFlow.js expands the accessibility of machine learning,
highlighting the broader impact and applications of CNNs beyond traditional
frameworks.
7. These references collectively validate the conclusions drawn in the document,
providing a robust foundation of scholarly support for the research findings.