# FCN8s-ResNet50

**Repository Path**: zevision/FCN8s-ResNet50

## Basic Information

- **Project Name**: FCN8s-ResNet50
- **Description**: About
本项目复现了 FCN-8s 网络，基于 PyTorch 实现，使用 ResNet-50 作为特征提取骨干。
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 3
- **Forks**: 0
- **Created**: 2025-09-09
- **Last Updated**: 2026-05-27

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

<div align="center">

# FCN8s-ResNet50

**Fully Convolutional Network (FCN) with ResNet-50 backbone for image semantic segmentation**

If you like this project, please give it a ⭐ support!

<p align="center">
    <img src="https://img.shields.io/github/last-commit/zevision/FCN8s-ResNet50">
    <img src='https://img.shields.io/github/languages/code-size/zevision/FCN8s-ResNet50.svg?style=flat-square' alt='Code size'>
    <img src='https://img.shields.io/github/commit-activity/t/zevision/FCN8s-ResNet50?color=red' alt='Last commit'>
</p>

</div>

---

## Project Introduction

This project reproduces the **FCN-8s** network based on **PyTorch**, using **ResNet-50** as the feature extraction backbone, focusing on **binary image semantic segmentation** tasks.

**FCN Paper**: [Fully Convolutional Networks for Semantic Segmentation (Long et al., 2015)](https://arxiv.org/abs/1411.4038)

---

## Project Structure

```text
📁 dataset/                 # Dataset processing related code
📁 model/                   # Network model definitions
📁 script/                  # Training scripts
📁 inference/               # Inference scripts
📁 utils/                   # Utility functions
📁 Portrait-dataset-2000/   # Binary classification dataset
```

## Features

* Binary classification segmentation task
* Uses ResNet-50 as the feature extraction backbone
* Implements the FCN-8s architecture with multi-level feature fusion for precise segmentation
* Provides a complete training and inference pipeline
* Supports model weight saving and loading
* Automatically computes dataset mean and standard deviation for normalization

---

## Usage

### 1️⃣ Dataset Preparation

* Dataset: [Portrait-dataset-2000-PaddlePaddle](https://aistudio.baidu.com/datasetdetail/220355)
* The dataset contains original images and corresponding mask images
* Mask image naming format: `{image_name}_matte.png`

### 2️⃣ Model Training

Run the training script:

```bash
python script/train.py
```

During training, the best model weights will be automatically saved to `checkpoints/best_model.pth`.

### 3️⃣ Model Inference

Use the trained model for image segmentation:

```bash
python inference/predict_image.py
```

The script will load the model weights, perform segmentation on the specified image, and save the result.

---

## Model Architecture

![Model Architecture](./FCN8s-RESNET50.png)

* Uses pretrained **ResNet-50** as the feature extractor
* Upsamples and fuses multi-level feature maps
* Uses skip connections to integrate multi-scale information
* Outputs binary segmentation results with the same size as the input image

---

## Project Highlights

* High accuracy for binary segmentation tasks
* Uses pretrained ResNet-50 to enhance feature extraction capability
* Automatic normalization for more stable training
* Complete training and validation workflow with checkpoint resume support
* Visualization of training accuracy and loss curves

```

```