自动驾驶场景图像分割(Unet)

本文使用matlab环境，测试了自动驾驶场景的图像分割任务。分割网络使用Unet。

一千张标注图像，最终训练精度达到 90%。

 

 数据准备

场景预标注数据下载地址：Semantic Segmentation for Self Driving Cars | Kaggle

数据下载可能需要翻墙，如果无法下载可移步 自动驾驶场景图像-分割标注数据.zip

数据标注了13个类别，每个类别具体名字未知，随便给个类名测试。

imagedir="E:\DLDatasets\kaggle\SelfDriving\dataA\dataA\CameraRGB";
labeldir="E:\DLDatasets\kaggle\SelfDriving\dataA\dataA\CameraSeg";

classNames=["c0","c1","c2","c3","c4","c5","c6","c7","c8","c9","c10","c11","c12"];
labelIds=0:12;

imds=imageDatastore(imagedir, "ReadFcn",@readImage);
pxds=pixelLabelDatastore(labeldir,classNames,labelIds,"ReadFcn",@readLabel);

ds=combine(imds,pxds);

图像原始大小为 600x800，首先自定义数据读取函数将其缩放到 96x128，另外标注图像是三通道，需要将其转成单通道图像。

function [data] = readImage(path)
    data=imread(path);
    data=imresize(data,[96 128]);
end

function [data] = readLabel(path)
    data=imread(path);
    data=data(:,:,1);
    data=imresize(data,[96 128]);
end

定义网络及参数

% 定义网络
imageSize = [96 128 3];
numClasses = 13;
lgraph = unetLayers(imageSize, numClasses);
% 训练参数
options = trainingOptions('adam', ...
    'InitialLearnRate',1e-3, ...   % 1e-4不收敛，改为1e-3
    'MaxEpochs',100, ...
    "MiniBatchSize", 8,...
    'VerboseFrequency',50, ...
    'Plots','training-progress');

 训练

net = trainNetwork(ds,lgraph,options)

最终训练情况

 测试

pic = imread("E:\DLDatasets\kaggle\SelfDriving\dataA\dataA\CameraRgb\02_00_120.png");
msk = imread("E:\DLDatasets\kaggle\SelfDriving\dataA\dataA\CameraSeg\02_00_120.png");
pic = imresize(pic, [96,128]);
msk = imresize(msk, [96,128]);
msk=msk(:,:,1);
out2 = predict(net,pic);

[h,w,~] = size(out2);
result = uint8(zeros(h,w));

for y = 1:h
    for x = 1:w
        [~,i]=max(out2(y,x,:));
        result(y,x)=i-1;
    end
end

result = result*20;

msk = msk*20;

subplot(1,3,1)
imshow(pic)
subplot(1,3,2)
imshow(result,'Colormap',jet(255))
subplot(1,3,3)
imshow(msk,'Colormap',jet(255))