CHAPTER 2: LITERATURE REVIEW

2.0 Introduction

This chapter presents the theoretical foundations and existing research pertinent to the
development of a fruit recognition system for visually impaired individuals. It synthesizes key
concepts, technological frameworks, and scholarly studies to contextualize the current project.
The review identifies critical gaps in assistive technologies, particularly in real-time fruit
identification, and establishes the rationale for the system’s design.

2.1 Theoretical Background

2.1.1 Concepts

Visually impaired individuals encounter significant barriers in object identification, a

fundamental aspect of independent living. According to global health data, approximately 2.2
billion people experience vision impairment, with 1 billion cases lacking adequate intervention
(World Health Organization, 2021). Fruit identification poses unique challenges due to
morphological diversity—variations in shape, color, texture, and ripeness stages complicate non-
visual assessment. Traditional assistive technologies prioritize navigation aids (e.g., white canes)
and text-to-speech systems, neglecting real-time object recognition in natural settings. This gap
exacerbates dependency, as 78% of visually impaired individuals report requiring assistance
during grocery shopping (American Foundation for the Blind, 2020).

Theoretical frameworks emphasize the need for user-centered design in assistive technologies,
where solutions must align with the sensory and cognitive needs of end-users (Lazar et al.,
2017). Key concepts include:

 Accessibility: Designing interfaces usable without visual input, leveraging auditory,

tactile, or gestural feedback.

 Real-Time Processing: Delivering instantaneous responses (<2 seconds) to facilitate

decision-making in dynamic environments.

 Contextual Awareness: Providing beyond-identification data (e.g., ripeness, nutritional

facts) to enhance autonomy.

2.1.2 Technology

The system leverages advances in computer vision and deep learning, specifically Convolutional
Neural Networks (CNNs), to achieve accurate fruit recognition. CNNs excel in image
classification tasks, with specialized models achieving >99% accuracy using custom datasets
(Dubey & Jalal, 2021). The MobileNet-SSD architecture, optimized for mobile devices, balances
speed and efficiency, processing 39 frames per second (FPS) on dedicated hardware (Wang et
al., 2021).

2.2 Review of Related Literature

2.2.1 Assistive Technologies for Visually Impaired Individuals

Existing assistive technologies predominantly address navigation and text recognition, leaving
object identification underexplored. Wearable devices like e-glasses and smart canes utilize
ultrasonic sensors for obstacle detection but lack object recognition capabilities (Dakopoulos &
Bourbakis, 2010). Screen readers and OCR applications convert text to speech, yet they cannot
identify physical objects. A survey by Khan et al. (2019) highlights that 65% of assistive tools
focus on mobility, while only 12% support object interaction. This imbalance perpetuates
dependency in daily tasks like grocery shopping.

Studies emphasize the socioeconomic impact of this gap. Vision impairment costs the global
economy US411billion annually in lost productivity, far exceeding the US25 billion needed to
address unmet vision care needs (World Health Organization, 2021). Technologies enabling
independence could alleviate this burden, yet adoption remains low due to poor real-world
applicability. For instance, VizWiz, a crowd-sourced image recognition app, suffers from delays
(up to 5 minutes), rendering it impractical for time-sensitive scenarios (Bigham et al., 2010).

2.2.2 Fruit Recognition Systems

Computer vision research demonstrates high accuracy in fruit classification under controlled
conditions. Dubey and Jalal (2021) achieved 99% accuracy in disease detection using deep
CNNs, while Peng et al. (2025) reported false positive/negative rates of 6.62%/5.03% in defect
recognition. However, these systems prioritize agricultural applications (e.g., sorting, quality
control) rather than accessibility.

Mobile-based fruit recognition tools face limitations in robustness. Systems using COCO
datasets (e.g., YOLOv3) achieve <60% accuracy in real-world settings due to environmental
variability (Lin et al., 2014). Lighting fluctuations, overlapping objects, and diverse backgrounds
degrade performance, with accuracy dropping 20–40% outside laboratories (Gallego et al.,
2021). Additionally, most solutions lack accessible interfaces. For example, "FruitID" uses visual
dashboards, excluding visually impaired users (Niranjani & Selvakumar, 2017).

2.2.3 Real-Time Object Recognition on Mobile Devices

Real-time processing is critical for usability, yet mobile deployment introduces challenges.
MobileNet-SSD achieves 39 FPS on GPUs but performance declines on consumer devices due to
battery and memory constraints (Wang et al., 2021). Optimization techniques like model
quantization and pruning reduce computational load but often compromise accuracy. Ibrahim et
al. (2021) note that 80% of mobile object recognition systems prioritize speed over precision,
leading to misclassification in complex scenes.

Energy efficiency remains a barrier. LiKamWa et al. (2013) found that continuous image sensing
drains smartphone batteries in <2 hours. Adaptive processing—adjusting frame rates based on
battery levels—can extend usability but requires sophisticated algorithms. Few systems
integrate these features, limiting practical deployment.

2.2.4 Accessibility in User Interface Design

Accessible interfaces must bypass visual interaction. Voice-controlled systems (e.g., Google
Lens) offer partial solutions but require stable internet and struggle with background noise
(Ross et al., 2019). Haptic feedback, effective for alerts, is underutilized in object recognition
tools. A study by Steinfeld et al. (2005) found that 70% of visually impaired users prefer
multimodal feedback (voice + haptics) for complex tasks.

Integration with existing accessibility frameworks is inconsistent. While Android’s TalkBack and
iOS’s VoiceOver support third-party apps, many vision-based tools fail to leverage these systems
(Kane et al., 2007). This fragmentation forces users to navigate multiple interfaces, increasing
cognitive load.

2.3 Research/Literature/Knowledge Gap

Four critical gaps emerge from the literature:

1. Limited Real-Time Feedback: Existing systems incur delays (>4 seconds), hindering
decision-making in dynamic environments (e.g., markets). No solution combines
MobileNet-SSD’s speed (>15 FPS on mobile) with fruit-specific accuracy.

2. Inadequate Accessibility: Interfaces rely on visual elements, excluding visually impaired

users. Haptic/voice integration remains underdeveloped.

3. Environmental Robustness: Accuracy declines 20–40% in real-world settings due to

lighting/background variability. Systems lack adaptive processing for these conditions.

4. Contextual Information Deficiency: Tools provide identification but lack ripeness,

quality, or nutritional data—crucial for informed choices.
The current project addresses these gaps by:

 Adaptation for fruit recognition via transfer learning, targeting >90% accuracy.

 Implementing adaptive processing to maintain performance across environments.

 Integrating a contextual database for accuracy in identification of each fruit.

This approach bridges computer vision advances with user-centered design, offering a template
for future assistive technologies.

References

 American Foundation for the Blind. (2020). Grocery Shopping Challenges for Visually
Impaired Individuals.

 Bigham, J. P., et al. (2010). VizWiz: Nearly real-time answers to visual questions.
Proceedings of the 23rd Annual ACM Symposium on User Interface Software and
Technology.

 Dakopoulos, D., & Bourbakis, N. G. (2010). Wearable obstacle avoidance electronic travel
aids for blind: A survey. IEEE Transactions on Systems, Man, and Cybernetics, Part C
(Applications and Reviews), 40(1), 25–35.

 Dubey, S. R., & Jalal, A. S. (2021). Fruit disease recognition using improved deep
convolutional neural network. Multimedia Tools and Applications, 80, 10295–10314.

 Gallego, A. J., et al. (2021). Fruit quality assessment using deep learning. Postharvest
Biology and Technology, 170.

 Ibrahim, M. S., et al. (2021). Real-time object detection on mobile devices: A review.
Journal of Real-Time Image Processing, 18, 1305–1320.

 Kane, S. K., et al. (2007). Freedom to roam: A study of mobile device adoption and
accessibility for people with visual and motor disabilities. Proceedings of the 9th
International ACM SIGACCESS Conference on Computers and Accessibility.

 Khan, A. R., et al. (2019). Assistive technologies for visually impaired people: A survey.
2019 International Conference on Electrical, Communication, and Computer Engineering
(ICECCE).

 Lin, T. Y., et al. (2014). Microsoft COCO: Common Objects in Context. European
Conference on Computer Vision.
 LiKamWa, R., et al. (2013). Energy characterization and optimization of image sensing
toward continuous mobile vision. Proceedings of the 11th Annual International
Conference on Mobile Systems, Applications, and Services.

 Niranjani, S. S., & Selvakumar, R. K. (2017). Fruit recognition using image features and
machine learning. 2017 International Conference on Intelligent Computing and Control
Systems (ICICCS).

 Peng, C. M., et al. (2025). Application of deep learning for fruit defect recognition in
Psidium guajava L. Scientific Reports, 15(1).

 Ross, D. A., et al. (2019). Accessible user interface design for visually impaired users.
ACM Transactions on Accessible Computing (TACCESS), 12(1), 1–32.

 Steinfeld, C. E., et al. (2005). Accessible interfaces for wearable computers. Proceedings
of the 7th International ACM SIGACCESS Conference on Computers and Accessibility.

 Wang, L., et al. (2021). An Improved Mobilenet-SSD Approach For Face Detection. 2021
IEEE International Conference on Computer Science and Artificial Intelligence (CSAI).

 World Health Organization. (2021). Blindness and vision impairment. Fact sheet.