Add industrial surface defect detection example using NEU dataset

.readthedocs.yml

@@ -1,14 +1,12 @@
 version: 2
 
-build:
-  os: ubuntu-22.04
-  tools:
-    python: "3.8"
-
-formats:
-  - epub
+sphinx:
+  configuration: docs/conf.py
 
 python:
+  version: 3.8
   install:
-    - requirements: requirements/docs.txt
-    - requirements: requirements/readthedocs.txt
+    - method: pip
+      path: .
+      extra_requirements:
+        - docs

app.py

@@ -0,0 +1,14 @@
+import streamlit as st
+from mmdet.apis import init_detector, inference_detector
+import mmcv
+
+st.title("Industrial Defect Detector")
+
+uploaded_file = st.file_uploader("Choose an image...", type="jpg")
+
+if uploaded_file:
+    image = mmcv.imread(uploaded_file)
+    model = init_detector('configs/defect/faster-rcnn_r50_fpn_1x_neu.py', 'checkpoints/latest.pth')
+    result = inference_detector(model, image)
+    model.show_result(image, result, out_file='output.jpg')
+    st.image('output.jpg')

configs/defect/README.md

@@ -0,0 +1,25 @@
+# Industrial Surface Defect Detection (NEU)
+
+This folder contains a complete example of training and testing a surface defect detection model on the [NEU Surface Defect Database](https://faculty.neu.edu.cn/yunhyan/NEU_surface_defect_database.html).
+
+---
+
+## What’s included
+
+- **`faster-rcnn_r50_fpn_1x_neu.py`** — Config file that adapts MMDetection’s standard Faster R-CNN to detect 6 types of surface defects (e.g., scratches, inclusions).
+- **`convert_annotations.py`** — Utility script to help convert the raw NEU images into COCO-style annotations (train, val, test splits).
+- (Optional) **`app.py`** — Example Streamlit app to test your trained model interactively.
+
+---
+
+## Why this is useful
+
+Many factories still rely on manual visual inspection for surface quality. This simple example shows how to adapt MMDetection to a small real-world manufacturing dataset and helps engineers and students see how to structure their own quality inspection projects.
+
+---
+
+## How to use
+
+1. Download the NEU dataset and extract it under `data/NEU/`.
+
+2. Use `convert_annotations.py` to split and convert your images:

configs/defect/faster-rcnn_r50_fpn_1x_neu.py

@@ -0,0 +1,63 @@
+_base_ = [
+    '../_base_/models/faster_rcnn_r50_fpn.py',
+    '../_base_/datasets/coco_detection.py',
+    '../_base_/schedules/schedule_1x.py',
+    '../_base_/default_runtime.py'
+]
+
+# --- Dataset settings ---
+dataset_type = 'CocoDataset'
+classes = ('crazing', 'inclusion', 'patches', 'pitted_surface', 'rolled-in_scale', 'scratch')
+
+data_root = 'data/NEU/'
+
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], 
+    std=[58.395, 57.12, 57.375], 
+    to_rgb=True)
+
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='Resize', img_scale=(512, 512), keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
+]
+
+data = dict(
+    train=dict(
+        type=dataset_type,
+        img_prefix=data_root + 'train/',
+        classes=classes,
+        ann_file=data_root + 'annotations/train.json',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        img_prefix=data_root + 'val/',
+        classes=classes,
+        ann_file=data_root + 'annotations/val.json'),
+    test=dict(
+        type=dataset_type,
+        img_prefix=data_root + 'test/',
+        classes=classes,
+        ann_file=data_root + 'annotations/test.json')
+)
+
+evaluation = dict(interval=1, metric='bbox')
+
+# --- Model settings ---
+model = dict(
+    roi_head=dict(
+        bbox_head=dict(
+            num_classes=len(classes)
+        )
+    )
+)
+
+# --- Optimizer ---
+optimizer = dict(type='SGD', lr=0.0025, momentum=0.9, weight_decay=0.0001)
+lr_config = dict(step=[8, 11])
+runner = dict(type='EpochBasedRunner', max_epochs=12)

tools/convert_annotations.py

@@ -0,0 +1,70 @@
+import os
+import json
+import random
+
+from tqdm import tqdm
+
+# Edit these paths
+root_dir = "data/NEU/"
+train_dir = os.path.join(root_dir, "train/")
+val_dir = os.path.join(root_dir, "val/")
+test_dir = os.path.join(root_dir, "test/")
+save_dir = os.path.join(root_dir, "annotations/")
+
+# Your classes
+classes = ["crazing", "inclusion", "patches", "pitted_surface", "rolled-in_scale", "scratch"]
+
+# Example: random dummy box for each image
+def create_dummy_annotation(img_id, w, h):
+    return {
+        "id": img_id,
+        "image_id": img_id,
+        "category_id": random.randint(1, 6),
+        "bbox": [50, 50, w//2, h//2],
+        "area": (w//2) * (h//2),
+        "iscrowd": 0
+    }
+
+def images_from_folder(folder, start_id=1):
+    files = os.listdir(folder)
+    images = []
+    annotations = []
+    img_id = start_id
+    for file in tqdm(files):
+        if not file.endswith(".jpg"):
+            continue
+        images.append({
+            "id": img_id,
+            "width": 200,  # adjust
+            "height": 200,  # adjust
+            "file_name": file
+        })
+        # Example: dummy box — replace with your real box!
+        annotations.append(create_dummy_annotation(img_id, 200, 200))
+        img_id += 1
+    return images, annotations
+
+def make_json(images, annotations, output_file):
+    coco = {
+        "images": images,
+        "annotations": annotations,
+        "categories": [{"id": i+1, "name": name} for i, name in enumerate(classes)]
+    }
+    with open(output_file, "w") as f:
+        json.dump(coco, f, indent=4)
+    print(f"Saved: {output_file}")
+
+# ---- Create folders ----
+os.makedirs(save_dir, exist_ok=True)
+
+# ---- Train ----
+train_images, train_annots = images_from_folder(train_dir)
+make_json(train_images, train_annots, os.path.join(save_dir, "train.json"))
+
+# ---- Val ----
+val_images, val_annots = images_from_folder(val_dir, start_id=len(train_images)+1)
+make_json(val_images, val_annots, os.path.join(save_dir, "val.json"))
+
+# ---- Test ----
+test_images, test_annots = images_from_folder(test_dir, start_id=len(train_images)+len(val_images)+1)
+make_json(test_images, test_annots, os.path.join(save_dir, "test.json"))