A New Lightweight Hybrid Model for Pistachio Classification

Project Code :TCMAPY2097

Objective

This project presents a lightweight hybrid model for classifying two types of pistachios: Kirmizi and Siirt. The goal is to achieve high accuracy while keeping computational costs low. EfficientNet-B0 serves as a baseline, while hybrid models include MobileNet with Attention, ResNet18 with SE, and ResNet18 with CBAM to enhance feature extraction. Attention mechanisms improve channel-wise and spatial sensitivity without significantly increasing model size. The dataset contains structured, class-specific features that support learning of shape and statistical patterns. A Flask-based web application enables users to register, log in, upload data, and receive predictions. The study emphasizes balancing accuracy, speed, and model efficiency. Overall, the project integrates efficient model design with an accessible user interface for testing.

Abstract

This project presents a lightweight hybrid model for pistachio classification using two classes: Kirmizi Pistachio and Siirt Pistachio. The aim is to design models that achieve strong performance while keeping computation low. Several architectures are explored, including EfficientNet-B0 as a simple baseline, and three hybrid variants: MobileNet with Attention, ResNet18 with Squeeze-and-Excitation (SE), and ResNet18 with Convolutional Block Attention Module (CBAM). These architectures are chosen to enhance feature extraction while maintaining a compact structure suitable for limited-resource environments.
The dataset includes class-specific features, allowing the models to learn shape-based and statistical patterns. The system is implemented using Flask as the backend, with HTML, CSS, and JavaScript for user interaction. The application includes modules for registration, login, classification, and logout, ensuring a complete interface for testing the models. The study focuses on achieving balanced accuracy, reduced complexity, and efficient deployment. Results show that hybrid models with attention provide noticeable improvements without increasing model size significantly. This approach demonstrates how attention modules can boost feature sensitivity while preserving fast inference.

Keywords: pistachio classification, lightweight models, EfficientNet-B0, MobileNet, ResNet18, SE module, CBAM, attention mechanism, feature extraction, Flask application

NOTE: Without the concern of our team, please don't submit to the college. This Abstract varies based on student requirements.