BUBT

Bangladesh University of Business and Technology

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING (CSE)
“Exploring Multiclass Fruits Recognition:
A Deep Learning Approch with XAI Insights”
PROJECT PROPOSAL
COURSE CODE : CSE 498A
COURSE TITLE : CAPSTONE PROJECT

SUBMITTED BY SUBMITTED TO
Md Darun Nayeem Mr. Sudipto Chaki
ID: 20211103029 Assistant Professor
INTAKE: 46th SECTION: 01 Department of Computer Science and
Engineering (CSE)
Md Sadikujjaman
ID: 19201103132
INTAKE: 43th SECTION: 04

Feroz Ahmed
ID: 20211103026
INTAKE: 46th SECTION: 01

Md Azizur Rahman PROJECT SUPERVISOR

ID: 20211103014 Sondip Poul Singh
INTAKE: 46th SECTION: 01 Assistant Professor
Department of Computer Science and
Zannatul Ferdous Fifa Engineering (CSE)
ID: 20211103028
INTAKE: 46th SECTION: 02
Exploring Multiclass Fruits Recognition: A Deep Learning Approch
with XAI Insights

MOTIVATION:
Fruit is in great demand by many people, and the classification of fruits is a
challenging task because of the different similarity of the fruit. A system can take
the traditional way of classifying fruits, but it is time-consuming and results in
lower productivity. Fruit is now widely sold in various place, such as supermarkets,
markets, malls, and others. However, sometimes it is difficult to distinguish the
types of fruit. The classification system helps customers to determine the kind of
fruit. Researchers have proposed several methods to build classification models for
fruits, the leading technology used in fruit recognition and yield calculation is the
intelligent detection system, in which image techniques can be applied, but Most of
the literature available right now does not provide enough accuracy for them to be
to have practical implementations. So more research is required. Convolutional
neural network (CNN) is the most commonly used approach in classification
problem. However, they have largely been used as black-box predictors, lacking
explanation for the underlying reasons. Explainable artificial intelligence (XAI) is
an emerging subfield of AI seeking to understand how models make their decisions.
In this work, we will apply XAI visualization to gain an insight into the features
learned by our CNN model.

OBJECTIVES:
Most approaches to fruits prediction at the moment has an average accuracy rate of
95%. Our principal objective is to find a method to predict individual fruit that is
much more accurate than the methods currently available in relevant literature.
Other objectives of ours include –
 Using CNN hybrid models (Including pre-trained models) to find best
accuracy .
 Trying Explainable AI (XAI) to insights model predictions & making it more
understandable.
 Developed a web interface for the users.
CRITICAL CHALLENGES:
 Developing an effective Convolutional neural network method to bridge the
species gap in deep learning models for fruit detection and disease
classification.
 Overcoming the limitations of unsupervised learning approaches to reduce
the high labeling cost associated with diverse fruit species.
 CNNs are highly complex models with numerous layers and millions of
parameters. Interpreting the decisions made by such models becomes
challenging due to this complexity.
 Black Box Nature: CNNs are often perceived as "black box" models because
they lack transparency in how they arrive at their decisions. Understanding
the rationale behind the model's predictions is crucial for trust and
accountability, especially in critical applications.
 Feature Attribution: Identifying which features or parts of the input data
contribute most to the model's decision-making process is not straightforward
in CNNs. Techniques for feature attribution need to be developed to provide
meaningful explanations.

CONFLICTING REQUIREMENTS:
Addressing Complex Engineering Problems (Ps) Through This Project –
 P1: Depth of knowledge required
o The project requires knowledge of deep learning and explainable AI (K8)
o Engineering design and development (K5 & K6)
o Knowledge of deep learning (K3) and explainable AI (K4)

 P2: Range of conflicting requirements

Huge fruits image data are required, which is very hard to collect or select the
proper and perfect data set for classification and training. It is difficult to get
more accuracy through tuning the CNN model. XAI is also a challenge to
handle.

 P3: Depth of analysis required

Select data set and preprocess the data. Apply CNN models on the data to get
predictions and then apply XAI to get the explanation of how the CNN model
predicts.
  P4: Familiarity of issues
CNN models are available, no issues to training the data set and get
predictions. XAI algorithm is not too much available, so there could be
problem while explaining the model.

 P5: Extent of applicable codes

LRP (XAI) method will be used for our CNN model. If it not works then we
will use LIME or SHAP method.

 P6: Extent of stakeholder involvement and conflicting requirements

o Kaggle notebook for data set.
o CampusX youtube chanel for CNN.
o Git hub code for XAI.

 P7: Interdependence
o Fruit data set selection.
o CNN model structure and train the data set.
o LRP ( explanation CNN ).

Addressing Complex Engineering Activities (As) Through This Project –

 A1: Range of resources
o Image data selection
o Data augmentation
o H5 or pickle model saving
o XAI visualization

 A3: Innovation
o Tuning the CNN model for better accuracy.
o XAI for CNN model to know how the model predict and why it predict.

 A5: Familiarity
The research deals with image classification and explanation about the deep
convolutional networks models.

Dataset Link: https://www.kaggle.com/datasets/moltean/fruits