PaddleRS：使用超分模块提高真实的低分辨率无人机影像的分割精度

%cd /home/aistudio/import matplotlib.pyplot as pltfrom PIL import Image

output = Image.open(r"work/example/Seg/UDD6_result/added_prediction/000161.JPG")

plt.figure(figsize=(18,12))#设置窗口大小plt.imshow(output), plt.axis('off')

img = Image.open(r"work/example/Seg/gt_result/data_05_2_14.png")
lq = Image.open(r"work/example/Seg/lq_result/added_prediction/data_05_2_14.png")
sr = Image.open(r"work/example/Seg/sr_result/added_prediction/data_05_2_14.png")

plt.figure(figsize=(18, 12))
plt.subplot(1,3,1), plt.title('GT')
plt.imshow(img), plt.axis('off')
plt.subplot(1,3,2), plt.title('predict_LR')
plt.imshow(lq), plt.axis('off')
plt.subplot(1,3,3), plt.title('predict_SR')
plt.imshow(sr), plt.axis('off')
plt.show()

add_lb = Image.open(r"work/example/Seg/gt_result/data_05_2_19.png")
lb = Image.open(r"work/example/Seg/gt_label/data_05_2_19.png")
img = Image.open(r"work/ValData/DJI300/data_05_2_19.png")

plt.figure(figsize=(18, 12))
plt.subplot(1,3,1), plt.title('image')
plt.imshow(img), plt.axis('off')
plt.subplot(1,3,2), plt.title('label')
plt.imshow(lb), plt.axis('off')
plt.subplot(1,3,3), plt.title('add_label')
plt.imshow(add_lb), plt.axis('off')
plt.show()

# 从github上克隆仓库!git clone https://github.com/PaddleCV-SIG/PaddleRS.git

# 安装依赖，大概一分多钟%cd PaddleRS/
!pip install -r requirements.txt

# 进行图像超分处理，使用的模型为DRNimport osimport paddleimport numpy as npfrom PIL import Imagefrom paddlers.models.ppgan.apps.drn_predictor import DRNPredictor# 输出预测结果的文件夹output = r'../work/example' # 待输入的低分辨率影像位置input_dir = r"../work/ValData/DJI300" paddle.device.set_device("gpu:0") # 若是cpu环境，则替换为 paddle.device.set_device("cpu")predictor = DRNPredictor(output) # 实例化filenames = [f for f in os.listdir(input_dir) if f.endswith('.png')]for filename in filenames:
    imgPath = os.path.join(input_dir, filename)   
    predictor.run(imgPath) # 预测

 # 可视化import osimport matplotlib.pyplot as plt
%matplotlib inline

lq_dir = r"../work/ValData/DJI300" #低分辨率影像文件夹sr_dir = r"../work/example/DRN" #超分辨率影像所在文件夹img_list = [f for f in os.listdir(lq_dir) if f.endswith('.png')]
show_num = 3 # 展示多少对影像for i in range(show_num):
    lq_box = (100, 100, 175, 175)
    sr_box = (400, 400, 700, 700)
    filename = img_list[i]
    image = Image.open(os.path.join(lq_dir, filename)).crop(lq_box) # 读取低分辨率影像
    sr_img = Image.open(os.path.join(sr_dir, filename)).crop(sr_box) # 读取超分辨率影像

    plt.figure(figsize=(12, 8))
    plt.subplot(1,2,1), plt.title('Input')
    plt.imshow(image), plt.axis('off')
    plt.subplot(1,2,2), plt.title('Output')
    plt.imshow(sr_img), plt.axis('off')
    plt.show()

%cd ..# clone PaddleSeg的项目!git clone https://gitee.com/paddlepaddle/PaddleSeg

# 安装依赖%cd /home/aistudio/PaddleSeg
!pip install  -r requirements.txt

# 对低分辨率的无人机影像进行预测!python predict.py \
       --config ../work/segformer_b3_UDD.yml \
       --model_path ../work/best_model/model.pdparams \
       --image_path ../work/ValData/DJI300 \
       --save_dir ../work/example/Seg/lq_result

# 对使用DRN超分重建后的影像进行预测!python predict.py \
       --config ../work/segformer_b3_UDD.yml \
       --model_path ../work/best_model/model.pdparams \
       --image_path ../work/example/DRN \
       --save_dir ../work/example/Seg/sr_result

# 展示部分预测的结果%cd /home/aistudio/import matplotlib.pyplot as pltfrom PIL import Imageimport os

img_dir = r"work/example/Seg/gt_result" #低分辨率影像文件夹lq_dir = r"work/example/Seg/lq_result/added_prediction"sr_dir = r"work/example/Seg/sr_result/added_prediction" # 超分辨率预测的结果影像所在文件夹img_list = [f for f in os.listdir(img_dir) if f.endswith('.png') ]for filename in img_list:
    img = Image.open(os.path.join(img_dir, filename))
    lq_pred = Image.open(os.path.join(lq_dir, filename))
    sr_pred = Image.open(os.path.join(sr_dir, filename))

    plt.figure(figsize=(12, 8))
    plt.subplot(1,3,1), plt.title('GT')
    plt.imshow(img), plt.axis('off')
    plt.subplot(1,3,2), plt.title('LR_pred') 
    plt.imshow(lq_pred), plt.axis('off')
    plt.subplot(1,3,3), plt.title('SR_pred') 
    plt.imshow(sr_pred), plt.axis('off')
    plt.show()