Image Classification Using Multiple Convolutional Neural Networks on the Fashion-MNIST Dataset

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• The increasing elderly population necessitates scalable care solutions.
• Service robotics offers potential for automated assistance in household and elder care settings.
• Robotic clothing manipulation is a complex daily activity requiring accurate detection and classification.

Purpose of the Study:

• To enhance apparel image classification accuracy for robotic applications.
• To evaluate multiple neural network models for fashion image recognition.
• To investigate the performance of a novel Multiple Convolutional Neural Network (MCNN15).

Main Methods:

• Developed and tested four distinct neural network models for fashion image classification.
• Utilized the Fashion-MNIST, Fashion-Product, and a custom household dataset.
• Implemented a Multiple Convolutional Neural Network with 15 convolutional layers (MCNN15).

Main Results:

• The MCNN15 model achieved a state-of-the-art classification accuracy of 94.04% on the Fashion-MNIST dataset.
• MCNN15 demonstrated 60% accuracy on the Fashion-Product dataset.
• MCNN15 achieved 40% accuracy on the custom household dataset.

Conclusions:

• The MCNN15 model significantly improves apparel image classification accuracy, outperforming existing literature on Fashion-MNIST.
• Advanced image classification is crucial for enabling service robots in domestic environments.
• Further research is needed to optimize performance on more complex, real-world datasets for household robotics.