关于matconvnet的几个小程序（3）识别车牌字符

首先新建cnn_plate_setup_data.m文件，读取相应的图片和标签，并减去均值。

function imdb =cnn_plate_setup_data(datadir)
inputSize =[20,20,1];%输入图片的尺寸，最后的1代表图像为灰度图
subdir=dir(datadir);
imdb.images.data=[];
imdb.images.labels=[];
imdb.images.set = [] ;
imdb.meta.sets = {'train', 'val', 'test'} ;
image_counter=0;%对图片进行计数
trainratio=0.8;%0.8的图片用于训练
for i=3:length(subdir)
imgfiles=dir(fullfile(datadir,subdir(i).name));
imgpercategory_count=length(imgfiles)-2;
disp([i-2 imgpercategory_count]);%显示类名和类中的文件个数
image_counter=image_counter+imgpercategory_count;%计数现在的图像个数
for j=3:length(imgfiles)%对于每一类中的每一张图片
img=imread(fullfile(datadir,subdir(i).name,imgfiles(j).name));%读取
img=imresize(img, inputSize(1:2));%重新调整大小
img=single(img);
% [~,~,d]=size(img);
% if d==3
% img=rgb2gray(img);%如果是彩色图像，就要把彩色图像转换为灰度图像
% continue;
% end
imdb.images.data(:,:,:,end+1)=single(img);%存储在最后一个维度里
imdb.images.labels(end+1)= i-2;
if j-2<imgpercategory_count*trainratio
imdb.images.set(end+1)=1;
else
imdb.images.set(end+1)=3;%为甚么是3而不是2？和前面的苹果图像识别程序很相似···
end
end
end
dataMean=mean(imdb.images.data,4);
imdb.images.data = single(bsxfun(@minus,imdb.images.data, dataMean)) ;
imdb.images.data_mean = dataMean;
end

还需要设计网络的结构，在cnn_plate_init.m文件里实现：

function net =cnn_plate_init()
rng('default');
rng(0) ;
f=1/100 ;
net.layers = {};
net.layers{end+1} = struct('type', 'conv', ...
'weights', {{f*randn(3,3,1,20, 'single'), zeros(1, 20, 'single')}}, ...
'stride', 1, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'pool', ...
'method', 'max', ...
'pool', [2 2], ...
'stride', 2, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'relu') ;
net.layers{end+1} = struct('type', 'conv', ...
'weights', {{f*randn(3,3,20,100, 'single'),zeros(1,100,'single')}}, ...
'stride', 1, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'pool', ...
'method', 'max', ...
'pool', [2 2], ...
'stride', 2, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'relu') ;
net.layers{end+1} = struct('type', 'conv', ...
'weights', {{f*randn(3,3,100,34, 'single'),zeros(1,34,'single')}}, ...
'stride', 1, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'softmaxloss') ;
% Meta parameters
net.meta.inputSize = [20 20 1] ;
net.meta.trainOpts.learningRate = logspace(-3, -5, 100);
net.meta.trainOpts.numEpochs = 50 ;
net.meta.trainOpts.batchSize = 1000 ;
% Fill in defaul values
net = vl_simplenn_tidy(net) ;
end

最后是把这些串联起来，进行训练：

function [net, info] = cnn_plate()
run(fullfile(fileparts(mfilename('fullpath')),...
'..', '..', 'matlab', 'vl_setupnn.m')) ;
datadir='F:\resources\train\数字与字母车牌';
opts.expDir = fullfile(vl_rootnn, 'data', 'plate-baseline') ;
opts.imdbPath = fullfile(opts.expDir, 'imdb.mat');
if exist(opts.imdbPath,'file')
imdb=load(opts.imdbPath);
else
imdb=cnn_plate_setup_data(datadir);
mkdir(opts.expDir) ;
save(opts.imdbPath, '-struct', 'imdb') ;
end
net=cnn_plate_init();
net.meta.normalization.averageImage =imdb.images.data_mean ;
opts.train.gpus=[];
[net, info] = cnn_train(net, imdb, getBatch(opts), ...
'expDir', opts.expDir, ...
net.meta.trainOpts, ...
opts.train, ...
'val', find(imdb.images.set == 3)) ;%注意这里！！
function fn = getBatch(opts)
% --------------------------------------------------------------------
fn = @(x,y) getSimpleNNBatch(x,y) ;
end
function [images, labels] = getSimpleNNBatch(imdb, batch)
images = imdb.images.data(:,:,:,batch) ;
labels = imdb.images.labels(1,batch) ;
if opts.train.gpus > 0
images = gpuArray(images) ;
end
end
end

训练好之后写一个识别的demo：

run(fullfile(fileparts(mfilename('fullpath')),...
'..', '..', 'matlab', 'vl_setupnn.m')) ;
addpath ../../data/plate-baseline;
datadir='F:\resources\train\数字与字母车牌';
class=2;index=1;
subdir=dir(datadir);
imgfiles=dir(fullfile(datadir,subdir(class+2).name));
img=imread(fullfile(datadir,subdir(class+2).name,imgfiles(index+2).name));
imshow(img);
net=load('net-epoch-10.mat');
net=net.net;
im_=single(img);
im_=imresize(im_,net.meta.inputSize(1:2));
im_=im_ - net.meta.normalization.averageImage;
opts.batchNormalization = false ;
net.layers{end}.type = 'softmax';
res=vl_simplenn(net,im_);
scores=squeeze(gather(res(end).x));
[bestScore,best]=max(scores);
disp(subdir(best+2).name);
disp(bestScore);