一、X光安检图像检测挑战赛3.0

地址：https://challenge.xfyun.cn/topic/info?type=Xray-2022

1.赛事背景

X光安检是目前在城市轨交、铁路、机场、物流业广泛使用的物检手段。使用人工智能技术，辅助一线安检员进行X光安检判图，可以有效降低因为安检员经验、能力或工作状态造成的错漏检问题。在实际场景中，因待检测物品的多样性、成像角度、重叠遮挡等问题，X光安检图像检测算法研究存在一定挑战。

2.赛事任务

本赛事的任务是：基于科大讯飞提供的真实X光安检图像集构建检测模型，对X光安检图像中的指定类别的物品进行检测。

3.评审规则

3.1 数据说明

此次比赛提供带标注的训练数据，即待检测物品在包裹中的X光图像及其标注文件。

本次比赛标注文件中的类别为8类，包括：

刀（knife）、剪刀（scissors）、打火机（lighter）、优盘（USBFlashDisk）、压力容器（pressure）、带喷嘴塑料瓶（plasticBottleWithaNozzle）、公章（seal）、电池（battery）。

待识别物品的X光成像示意图如图所示。

比赛提供的X光图像及其矩形框标注的文件按照数据来源存放在不同的文件夹中，图像文件采用jpg格式，标注文件采用xml格式，各字段含义参照voc数据集。voc各字段含义对应表为：

  • filename 文件名

  • size 图像尺寸

    • width 图像宽度

    • height 图像高度

    • depth 图像深度，一般为3表示是彩色图像

  • object 图像中的目标，可能有多个

    • name 该目标的标签名称

    • bndbox 该目标的标注框

      • xmin 该目标的左上角宽度方向坐标

      • ymin 该目标的左上角高度方向坐标

      • xmax 该目标的右下角宽度方向坐标

      • ymax 该目标的右下角高度方向坐标

3.2 评估指标

评测方式采用计算mAP（IoU = 0.5）的方式。

首先计算每个类的AP：

（1）根据预测框和标注框的IoU是否达到阈值0.5判断该预测框是真阳性还是假阳性；

（2）根据每个预测框的置信度进行从高到低排序；

（3）在不同置信度阈值下计算精确率和召回率，得到若干组PR值；

（4）绘制PR曲线并计算AP值。

然后计算mAP：把所有类的AP值求平均得到mAP。

3.3 作品提交要求

作品必须健康、合法、无任何不良信息及商业宣传行为，不违反任何中华人民共和国有关法律。须保证原创性，不侵犯任何第三方知识产权或其他权利；一经发现或经权利人指出，主办方将直接取消其参赛资格，科大讯飞保留赛事解释权。

选手需要提交json格式文件，详情见示例。其中，坐标值必须为大于0的正数且不能超过图像的宽高。

按照赛题数据页面2022gamedatasettest1.txt里面的顺序组织json

提交文件需按序排列，首先按图片顺序排列，然后按类别顺序排列，置信度顺序随意。

  • (a) 图片顺序，请按照图片编号顺序。

  • (b) 类别顺序，请参照下列顺序： {‘knife’: 1, ‘scissors’: 2, ‘lighter’: 3, ‘USBFlashDisk’: 4, ‘pressure’: 5, ‘plasticBottleWithaNozzle’: 6, ‘seal’: 7, ‘battery’: 8}

二、数据集处理

https://user-images.githubusercontent.com/53808988/135017525-4b59ee12-65b1-4d9e-b135-8e3cec1651e6.png

1.标注文件整理

# 解压缩数据集
!unzip -qoa data/data165820/round1.zip -d data/
!unzip -qoa data/data165820/round2_test.zip -d data/

!mv data/讯飞研究院-X光安检图像检测挑战赛2022公开数据 data/round1

!mkdir data/round1/train/XML

import glob
import shutil
import os

def move_xml(domain_name):
    xml_dir=f"data/round1/train/{domain_name}/XML"
    filenames=glob.glob(f'{xml_dir}/*.xml')
    print(f"{xml_dir}移动的XML共有{len(filenames)}个。")
    if len(filenames)==0:
        return 
    for filename  in filenames:
        file_path, shortname=os.path.split(filename)
        new_name=domain_name+'_'+shortname
        new_name=os.path.join("data/round1/train/XML", new_name)
        shutil.move(filename, new_name)
    shutil.rmtree(xml_dir)
    print("重命名并移动文件完成。")

move_xml('domain1') 
move_xml('domain2')    
move_xml('domain3')

data/round1/train/domain1/XML移动的XML共有1323个。
重命名并移动文件完成。
data/round1/train/domain2/XML移动的XML共有1383个。
重命名并移动文件完成。
data/round1/train/domain3/XML移动的XML共有1308个。
重命名并移动文件完成。

print(len(glob.glob('data/round1/train/XML/*.xml')))

4014

!mv data/round1/train/XML data/round1/train/Annotations

2.图片文件整理

import glob
import shutil
import os

def move_xml(domain_name):
    jpg_dir=f"data/round1/train/{domain_name}/"
    filenames=glob.glob(f'{jpg_dir}/*.jpg')
    print(f"{jpg_dir}移动的XML共有{len(filenames)}个。")
    if len(filenames)==0:
        return 
    for filename  in filenames:
        file_path, shortname=os.path.split(filename)
        new_name=domain_name+'_'+shortname
        new_name=os.path.join("data/round1/train/JPEGImages", new_name)
        shutil.move(filename, new_name)
    shutil.rmtree(jpg_dir)
    print("重命名并移动文件完成。")

!mkdir data/round1/train/JPEGImages

move_xml('domain1')
move_xml('domain2')
move_xml('domain3')

data/round1/train/domain1/移动的XML共有1323个。
重命名并移动文件完成。
data/round1/train/domain2/移动的XML共有1383个。
重命名并移动文件完成。
data/round1/train/domain3/移动的XML共有1308个。
重命名并移动文件完成。

print(len(glob.glob('data/round1/train/JPEGImages/*.jpg')))

4014

3.PaddleX 安装

!python -m pip install --upgrade -q pip --user
!pip install -q -U paddlex

!pip list|grep paddle

paddle-bfloat          0.1.7
paddlefsl              1.0.0
paddlehub              2.0.4
paddlenlp              2.1.1
paddlepaddle-gpu       2.3.1.post101
paddleslim             2.2.1
paddlex                2.1.0
tb-paddle              0.3.6

4.数据集划分

!paddlex --split_dataset --format VOC --dataset_dir data/round1/train/ --val_value 0.203

[32m[08-24 01:01:01 MainThread @logger.py:242][0m Argv: /opt/conda/envs/python35-paddle120-env/bin/paddlex --split_dataset --format VOC --dataset_dir data/round1/train/ --val_value 0.203
[0m[33m[08-24 01:01:01 MainThread @utils.py:79][0m [5m[33mWRN[0m paddlepaddle version: 2.3.1. The dynamic graph version of PARL is under development, not fully tested and supported
[0m/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/parl/remote/communication.py:38: DeprecationWarning: 'pyarrow.default_serialization_context' is deprecated as of 2.0.0 and will be removed in a future version. Use pickle or the pyarrow IPC functionality instead.
  context = pyarrow.default_serialization_context()
[0m/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import MutableMapping
[0m/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Iterable, Mapping
[0m/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Sized
[0m2022-08-24 01:01:02,135-WARNING: type object 'QuantizationTransformPass' has no attribute '_supported_quantizable_op_type'
[0m2022-08-24 01:01:02,135-WARNING: If you want to use training-aware and post-training quantization, please use Paddle >= 1.8.4 or develop version
[0m2022-08-24 01:01:03 [INFO]	Dataset split starts...[0m
[0m2022-08-24 01:01:04 [INFO]	Dataset split done.[0m
[0m2022-08-24 01:01:04 [INFO]	Train samples: 3200[0m
[0m2022-08-24 01:01:04 [INFO]	Eval samples: 814[0m
[0m2022-08-24 01:01:04 [INFO]	Test samples: 0[0m
[0m2022-08-24 01:01:04 [INFO]	Split files saved in data/round1/train/[0m
[0m[0m

!ls data/round1/train/

Annotations  JPEGImages  labels.txt  train_list.txt  val_list.txt

5.图片查看

在这里插入图片描述

在这里插入图片描述

在这里插入图片描述

三、模型训练

1.transforms定义

# 定义训练和验证时的transforms
# API说明：https://github.com/PaddlePaddle/PaddleX/blob/develop/docs/apis/transforms/transforms.md
import paddlex as pdx
from paddlex import transforms as T

train_transforms = T.Compose([
    # T.MixupImage(mixup_epoch=-1), 
    T.RandomDistort(),
    T.RandomHorizontalFlip(), 
    T.RandomVerticalFlip(),         
    T.BatchRandomResize(
        target_sizes=[320, 352, 384, 416, 448, 480, 512, 544, 576, 608],
        interp='RANDOM'), 
    # T.Resize(target_size=224, interp='LINEAR'),
    T.Normalize(
            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

eval_transforms = T.Compose([
    T.Resize(
        224, interp='CUBIC'), T.Normalize(
            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

[08-24 01:01:06 MainThread @utils.py:79] WRN paddlepaddle version: 2.3.1. The dynamic graph version of PARL is under development, not fully tested and supported


/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/parl/remote/communication.py:38: DeprecationWarning: 'pyarrow.default_serialization_context' is deprecated as of 2.0.0 and will be removed in a future version. Use pickle or the pyarrow IPC functionality instead.
  context = pyarrow.default_serialization_context()
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import MutableMapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Iterable, Mapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Sized
2022-08-24 01:01:07,469-WARNING: type object 'QuantizationTransformPass' has no attribute '_supported_quantizable_op_type'
2022-08-24 01:01:07,470-WARNING: If you want to use training-aware and post-training quantization, please use Paddle >= 1.8.4 or develop version

2.数据集定义

# 定义训练和验证所用的数据集
# API说明：https://github.com/PaddlePaddle/PaddleX/blob/develop/docs/apis/datasets.md
train_dataset = pdx.datasets.VOCDetection(
    data_dir='data/round1/train/',
    file_list='data/round1/train/train_list.txt',
    label_list='data/round1/train/labels.txt',
    transforms=train_transforms,
    shuffle=True)

eval_dataset = pdx.datasets.VOCDetection(
    data_dir='data/round1/train',
    file_list='data/round1/train/val_list.txt',
    label_list='data/round1/train/labels.txt',
    transforms=eval_transforms,
    shuffle=False)

2022-08-22 23:08:02 [INFO]	Starting to read file list from dataset...
2022-08-22 23:08:06 [INFO]	3200 samples in file data/round1/train/train_list.txt, including 3200 positive samples and 0 negative samples.
creating index...
index created!
2022-08-22 23:08:06 [INFO]	Starting to read file list from dataset...
2022-08-22 23:08:07 [INFO]	814 samples in file data/round1/train/val_list.txt, including 814 positive samples and 0 negative samples.
creating index...
index created!

3.模型定义

# YOLO检测模型的预置anchor生成
# API说明: https://github.com/PaddlePaddle/PaddleX/blob/release/2.0.0/paddlex/tools/anchor_clustering/yolo_cluster.py
import numpy as np
anchors = train_dataset.cluster_yolo_anchor(num_anchors=9, image_size=480)
anchor_masks = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]

# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标，参考https://github.com/PaddlePaddle/PaddleX/tree/release/2.0.0/tutorials/train#visualdl可视化训练指标
num_classes = len(train_dataset.labels)
model = pdx.det.YOLOv3(
    num_classes=num_classes,
    backbone='DarkNet53',
    anchors=anchors.tolist() if isinstance(anchors, np.ndarray) else anchors,
    anchor_masks=[[6, 7, 8], [3, 4, 5], [0, 1, 2]],
    label_smooth=True,
    ignore_threshold=0.6)

2022-08-22 23:09:01 [INFO]	Running kmeans for 9 anchors on 6829 points...


Evolving anchors with Genetic Algorithm: fitness = 0.7607: 100%|██████████| 1000/1000 [00:02<00:00, 493.54it/s]
W0822 23:09:04.492769    98 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 10.1
W0822 23:09:04.496325    98 gpu_resources.cc:91] device: 0, cuDNN Version: 7.6.

4.模型定义

主要是batch size 选择：

  • batch size: 8， 对应显存 13421
  • batch size:19.37， 对应显存 32000
    选择16即可。
Sun Aug 21 17:54:18 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.32.03    Driver Version: 460.32.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  On   | 00000000:05:00.0 Off |                    0 |
| N/A   38C    P0    62W / 300W |  28077MiB / 32510MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

# API说明：https://github.com/PaddlePaddle/PaddleX/blob/release/2.0.0/paddlex/cv/models/detector.py
# 各参数介绍与调整说明：https://paddlex.readthedocs.io/zh_CN/develop/appendix/parameters.html
model.train(
    num_epochs=200,  # 训练轮次
    train_dataset=train_dataset,  # 训练数据
    eval_dataset=eval_dataset,  # 验证数据
    train_batch_size=16,  # 批大小
    pretrain_weights='COCO',  # 预训练权重
    learning_rate=0.005 / 12,  # 学习率
    warmup_steps=500,  # 预热步数
    warmup_start_lr=0.0,  # 预热起始学习率
    save_interval_epochs=5,  # 每5个轮次保存一次，有验证数据时，自动评估
    lr_decay_epochs=[85, 135],  # step学习率衰减
    save_dir='output/yolov3_darknet53',  # 保存路径
    use_vdl=True)  # 其用visuadl进行可视化训练记录

5.提交后台执行

完成上述步骤即可生成版本、提交后台运行，很快得到对应的模型。可下载输出结果，含notebook代码、log日志、保存的模型等信息文件。
visualdl 本地训练可视化如下：

在这里插入图片描述

四、模型预测

1.加载训练保存的模型预测

模型在训练过程中，会每间隔一定轮数保存一次模型，在验证集上评估效果最好的一轮会保存在save_dir目录下的best_model文件夹。通过如下方式可加载模型，进行预测：

# 解压缩后台任务运行得到的最佳模型
!unzip -qoa data/data166176/best_model.zip

!wget https://ai-contest-static.xfyun.cn/2022/%E6%95%B0%E6%8D%AE%E9%9B%86/1/2022gamedatasettest1.txt
!wget https://ai-contest-static.xfyun.cn/2022/%E6%95%B0%E6%8D%AE%E9%9B%86/2022gamedatasettest2.txt

--2022-08-24 01:03:42--  https://ai-contest-static.xfyun.cn/2022/%E6%95%B0%E6%8D%AE%E9%9B%86/1/2022gamedatasettest1.txt
Resolving ai-contest-static.xfyun.cn (ai-contest-static.xfyun.cn)... 220.181.53.219
Connecting to ai-contest-static.xfyun.cn (ai-contest-static.xfyun.cn)|220.181.53.219|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 19057 (19K) [text/plain]
Saving to: ‘2022gamedatasettest1.txt.1’

2022gamedatasettest 100%[===================>]  18.61K  --.-KB/s    in 0s      

2022-08-24 01:03:43 (289 MB/s) - ‘2022gamedatasettest1.txt.1’ saved [19057/19057]

--2022-08-24 01:03:43--  https://ai-contest-static.xfyun.cn/2022/%E6%95%B0%E6%8D%AE%E9%9B%86/2022gamedatasettest2.txt
Resolving ai-contest-static.xfyun.cn (ai-contest-static.xfyun.cn)... 220.181.53.219
Connecting to ai-contest-static.xfyun.cn (ai-contest-static.xfyun.cn)|220.181.53.219|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 76209 (74K) [text/plain]
Saving to: ‘2022gamedatasettest2.txt.1’

2022gamedatasettest 100%[===================>]  74.42K  --.-KB/s    in 0.03s   

2022-08-24 01:03:43 (2.73 MB/s) - ‘2022gamedatasettest2.txt.1’ saved [76209/76209]

%cd ~

img_file=[]
with open('2022gamedatasettest1.txt','r') as f:
    img_file=f.readlines()
print(len(img_file))   

/home/aistudio
1003

%cd data/round1/test/test1

import paddlex as pdx
model = pdx.load_model('/home/aistudio/best_model')
result=[]
# 不要批量预测，显存容易爆
for img in img_file:
    item = model.predict(img.strip('\n'))
    result.append(item)

/home/aistudio/data/round1/test/test1


W0824 01:04:04.971199   145 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 10.1
W0824 01:04:04.978477   145 gpu_resources.cc:91] device: 0, cuDNN Version: 7.6.


2022-08-24 01:04:05 [INFO]	Model[YOLOv3] loaded.

for item in result[5]:
    print(item)

{'category_id': 2, 'category': 'knife', 'bbox': [97.13526916503906, 365.67645263671875, 114.161376953125, 209.08056640625], 'score': 0.020859738811850548}

label_list = ['knife','scissors', 'lighter', 'USBFlashDisk', 'pressure', 'plasticBottleWithaNozzle', 'seal', 'battery']

print(len(result[0]))
for item in result[0]:
    print(item)

6
{'category_id': 0, 'category': 'USBFlashDisk', 'bbox': [101.52272033691406, 393.29205322265625, 50.66070556640625, 80.339599609375], 'score': 0.01219225861132145}
{'category_id': 0, 'category': 'USBFlashDisk', 'bbox': [87.89407348632812, 260.2281494140625, 53.05218505859375, 102.7266845703125], 'score': 0.010960005223751068}
{'category_id': 1, 'category': 'battery', 'bbox': [109.76924133300781, 261.1977844238281, 38.988037109375, 63.71624755859375], 'score': 0.12186391651630402}
{'category_id': 1, 'category': 'battery', 'bbox': [81.07035827636719, 261.7168273925781, 58.88621520996094, 123.1175537109375], 'score': 0.09327761828899384}
{'category_id': 1, 'category': 'battery', 'bbox': [264.4331970214844, 305.2523193359375, 47.3126220703125, 64.0133056640625], 'score': 0.02472269907593727}
{'category_id': 1, 'category': 'battery', 'bbox': [101.52272033691406, 393.29205322265625, 50.66070556640625, 80.339599609375], 'score': 0.016887478530406952}

2.格式化提交结果

%cd ~

import os
import json

label_list = ['knife','scissors', 'lighter', 'USBFlashDisk', 'pressure', 'plasticBottleWithaNozzle', 'seal', 'battery']

save_result=[]
for item in result:
    temp_result = [[], [], [], [], [], [], [], []]
    for item2 in item:        
        if not len(item2):
            continue
        # name序号对齐
        name=item2['category']
        index = label_list.index(name)
        # 获取bbox
        bbox=item2['bbox']
        # 获取置信度
        score=item2['score']
        temp_result[index].append([bbox[0], bbox[1], bbox[0]+bbox[2], bbox[3]+bbox[1], score])        
    save_result.append(temp_result)

with open('/home/aistudio/result.json', 'w') as fp:
0], bbox[1], bbox[0]+bbox[2], bbox[3]+bbox[1], score])        
    save_result.append(temp_result)

with open('/home/aistudio/result.json', 'w') as fp:
    json.dump(save_result, fp, ensure_ascii=False)

/home/aistudio

此文章为搬运
原项目链接

# 计算机视觉 # 人工智能 # 深度学习
Logo
百度飞桨AI Studio社区

学大模型，用大模型上飞桨星河社区！每天8点V100G算力免费领！免费领取ERNIE 4.0 100w Token >>>

更多推荐

cover

利用Amazon Bedrock畅玩Claude 3等多种领先模型，抢占AI高地(体验倒计时4小时)

avatar 百度飞桨星河社区
cover

RAPTOR：索引树状 RAG，使用树结构来捕捉文本的高级和低级细节

avatar 百度飞桨星河社区

RAG介绍

LlamaIndex是一个基于大型语言模型（LLM）的应用程序数据框架，旨在连接LLM与外部数据，特别是私有或特定领域的数据，以实现更准确的文本生成和知识增强。

avatar 百度飞桨星河社区