Metadata-Version: 2.1
Name: minhtanfe
Version: 16.1.0
Summary: A high-performance, lightweight Feature Extractor with Alpha Gating for Edge AI.
Home-page: https://github.com/LeMinhTan/MinhTanFE
Author: Le Minh Tan
Author-email: tan.lmt.it@gmail.com
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.7
Description-Content-Type: text/markdown
Requires-Dist: torch>=1.7.0

Project Description: MinhTanFE V16.1

1. Overview

MinhTanFE (Minh Tan Feature Extractor) is a lightweight Convolutional Neural Network (CNN) architecture designed for optimized image feature recognition tasks on edge devices (Edge AI). Version V16.1 focuses on balancing extremely high accuracy with superior processing speed through an intelligent information filtering mechanism.

2. Core Architecture

This architecture is built upon three key technical pillars:

Alpha Gating (Spatial Attention):
Utilizes an auxiliary branch with a Sigmoid activation function to generate a weighted “mask,” enabling the model to focus on regions containing important information (such as handwriting strokes) while filtering out background noise from the very first layer.
Depthwise Separable Convolution:
Replaces traditional Conv2D layers with a combination of Depthwise and Pointwise convolutions. This technique significantly reduces the number of parameters and computational cost (FLOPs) while preserving strong feature extraction capability.
Residual Learning (Skip Connection):
Incorporates identity mappings (skip connections) to stabilize gradient flow, prevent vanishing gradients, and allow faster convergence during training.
3. Benchmark Summary

The model has been evaluated on three standard datasets, showing impressive results compared to a traditional 3-layer Conv baseline:

A-Z Handwritten: Achieved 95.70% accuracy (25% faster training speed).
EMNIST Letters: Achieved 92.70% accuracy (maintains stability on noisy data).
FashionMNIST: Achieved 87.39% accuracy (demonstrates generalization on non-character objects).
