DNN Image Classification

Sources:

function dnn_image_classification_demo(im, name, crop)
    % input image (BGR channel order)
    if nargin < 1 || isempty(im)
        im = fullfile(mexopencv.root(), 'test', 'space_shuttle.jpg');
    end
    img = cv.imread(im, 'Color',true, 'FlipChannels',false);

    % import pretrained model
    if nargin < 2, name = 'GoogLeNet'; end
    fprintf('Load model...   '); tic;
    switch lower(name)
        case 'alexnet'
            % ImageNet ILSVRC 2012
            [net, labels, blobOpts] = AlexNet();
        case 'caffenet'
            % ImageNet ILSVRC 2012
            [net, labels, blobOpts] = CaffeNet();
        case 'vggnet'
            % ImageNet ILSVRC 2012
            [net, labels, blobOpts] = VGGNet();
        case 'inception'
            % ImageNet ILSVRC 2012
            [net, labels, blobOpts] = Inception();
        case 'googlenet'
            % ImageNet ILSVRC 2012
            [net, labels, blobOpts] = GoogLeNet();
        case 'resnet'
            % ImageNet ILSVRC 2012
            [net, labels, blobOpts] = ResNet();
        case 'squeezenet'
            % ImageNet ILSVRC 2012
            [net, labels, blobOpts] = SqueezeNet('v1.1');
        otherwise
            error('Unrecognized model %s', name)
    end
    toc;
    assert(~net.empty(), 'Failed to read network %s', name);

    % feed image to network
    if nargin < 3, crop = true; end
    blobOpts = ['Crop',crop, blobOpts];
    blob = cv.Net.blobFromImages(img, blobOpts{:});
    net.setInput(blob);

    % run forward pass
    fprintf('Forward pass... '); tic;
    prob = net.forward();  % 1-by-nclasses-by-1-by-1
    toc;

    % prepare output image
    out = flip(img, 3);  % BGR to RGB

    % prediction: image classification and top-5 predictions
    [~,ord] = sort(prob, 'descend');
    disp('Top-5 predictions:')
    for i=1:5
        fprintf('%6.2f%% = %s\n', prob(ord(i))*100, labels{ord(i)});
    end
    imshow(out), title('Image Classification')
    xlabel(labels{ord(1)})
end

% --- Helper functions ---

function dname = get_dnn_dir(dname)
    %GET_DNN_DIR  Path to model files, and show where to get them if missing

    dname = fullfile(mexopencv.root(), 'test', 'dnn', dname);
    b = isdir(dname);
    if ~b
        % display help of calling function
        % (assumed to be a local function in current file)
        st = dbstack(1);
        help([mfilename() filemarker() st(1).name])
    end
    assert(b, 'Missing model: %s', dname);
end

function labels = readLabels(labelsFile, skipFirstWord)
    if nargin < 2, skipFirstWord = false; end
    if ~mexopencv.isOctave()
        fid = fopen(labelsFile, 'rt');
        C = textscan(fid, '%s', 'Delimiter','\n');
        fclose(fid);
        labels = C{1};
    else
        %HACK: textscan is buggy and unreliable in Octave!
        labels = textread(labelsFile, '%s', 'Delimiter','\n');
    end
    if skipFirstWord
        labels = regexprep(labels, '^\w+\s*', '', 'once');
    end
end

% --- Pretrained models ---
% See also: https://github.com/opencv/opencv_extra/blob/3.3.1/testdata/dnn/download_models.py

function [net, labels, blobOpts] = AlexNet()
    %ALEXNET  BAIR/BVLC AlexNet Model [Caffe]
    %
    % homepage = https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet
    %
    % ## Model
    %
    % file = test/dnn/AlexNet/deploy.prototxt
    % url  = https://github.com/BVLC/caffe/raw/master/models/bvlc_alexnet/deploy.prototxt
    % hash = cb77655eb4db32c9c47699c6050926f9e0fc476a
    %
    % ## Weights
    %
    % file = test/dnn/AlexNet/bvlc_alexnet.caffemodel
    % url  = http://dl.caffe.berkeleyvision.org/bvlc_alexnet.caffemodel
    % hash = 9116a64c0fbe4459d18f4bb6b56d647b63920377
    % size = 232 MB
    %
    % ## Classes
    %
    % file = test/dnn/AlexNet/synset_words.txt
    % url  = https://github.com/opencv/opencv/raw/3.3.1/samples/data/dnn/synset_words.txt
    %

    dname = get_dnn_dir('AlexNet');
    net = cv.Net('Caffe', ...
        fullfile(dname, 'deploy.prototxt'), ...
        fullfile(dname, 'bvlc_alexnet.caffemodel'));
    labels = readLabels(fullfile(dname, 'synset_words.txt'), true);
    blobOpts = {'SwapRB',false, 'Size',[227 227], 'Mean',[104 117 123]};
end

function [net, labels, blobOpts] = CaffeNet()
    %CAFFENET  BAIR/BVLC CaffeNet Model, a replication of AlexNet with some modification [Caffe]
    %
    % homepage = https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_caffenet
    %
    % ## Model
    %
    % file = test/dnn/CaffeNet/deploy.prototxt
    % url  = https://github.com/BVLC/caffe/raw/master/models/bvlc_reference_caffenet/deploy.prototxt
    % hash = 6a40fd8b77233afee8fd525ce59bb4ccaae78d58
    %
    % ## Weights
    %
    % file = test/dnn/CaffeNet/bvlc_reference_caffenet.caffemodel
    % url  = http://dl.caffe.berkeleyvision.org/bvlc_reference_caffenet.caffemodel
    % hash = 4c8d77deb20ea792f84eb5e6d0a11ca0a8660a46
    % size = 232 MB
    %
    % ## Classes
    %
    % file = test/dnn/CaffeNet/synset_words.txt
    % url  = https://github.com/opencv/opencv/raw/3.3.1/samples/data/dnn/synset_words.txt
    %

    dname = get_dnn_dir('CaffeNet');
    net = cv.Net('Caffe', ...
        fullfile(dname, 'deploy.prototxt'), ...
        fullfile(dname, 'bvlc_reference_caffenet.caffemodel'));
    labels = readLabels(fullfile(dname, 'synset_words.txt'), true);
    blobOpts = {'SwapRB',false, 'Size',[227 227], 'Mean',[104 117 123]};
end

function [net, labels, blobOpts] = VGGNet()
    %VGGNET  VGG team ILSVRC-2014 16-layer model [Caffe]
    %
    % homepage = http://www.robots.ox.ac.uk/~vgg/research/very_deep/
    %
    % ## Model
    %
    % file = test/dnn/VGGNet/VGG_ILSVRC_16_layers_deploy.prototxt
    % url  = https://gist.githubusercontent.com/ksimonyan/211839e770f7b538e2d8/raw/0067c9b32f60362c74f4c445a080beed06b07eb3/VGG_ILSVRC_16_layers_deploy.prototxt
    % hash = 2734e5500f1445bd7c9fee540c99f522485247bd
    %
    % ## Weights
    %
    % file = test/dnn/VGGNet/VGG_ILSVRC_16_layers.caffemodel
    % url  = http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel
    % hash = 9363e1f6d65f7dba68c4f27a1e62105cdf6c4e24
    % size = 527 MB
    %
    % ## Classes
    %
    % file = test/dnn/VGGNet/synset_words.txt
    % url  = https://github.com/opencv/opencv/raw/3.3.1/samples/data/dnn/synset_words.txt
    %

    dname = get_dnn_dir('VGGNet');
    net = cv.Net('Caffe', ...
        fullfile(dname, 'VGG_ILSVRC_16_layers_deploy.prototxt'), ...
        fullfile(dname, 'VGG_ILSVRC_16_layers.caffemodel'));
    labels = readLabels(fullfile(dname, 'synset_words.txt'), true);
    blobOpts = {'SwapRB',false, 'Size',[224 224], 'Mean',[103.939 116.779 123.68]};
end

function [net, labels, blobOpts] = Inception()
    %INCEPTION  Google Inception V1 model [TensorFlow]
    %
    % homepage = https://github.com/tensorflow/tensorflow
    %
    % ## Model + Weights + Classes
    %
    % file = test/dnn/Inception/tensorflow_inception_graph.pb
    % file = test/dnn/Inception/imagenet_comp_graph_label_strings.txt
    % url  = https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip
    % hash = e3b84c7e240ce8025b30d868f5e840b4bba9761d
    % size = 47.6 MB
    %

    dname = get_dnn_dir('Inception');
    net = cv.Net('Tensorflow', ...
        fullfile(dname, 'tensorflow_inception_graph.pb'));
    labels = readLabels(fullfile(dname, 'imagenet_comp_graph_label_strings.txt'), false);
    blobOpts = {'SwapRB',true, 'Size',[224 224], 'Mean',[117 117 117]};
end

function [net, labels, blobOpts] = GoogLeNet()
    %GOOGLENET  BAIR/BVLC GoogleNet Model, a Caffe implementation of Inception model [Caffe]
    %
    % homepage = https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet
    %
    % ## Model
    %
    % file = test/dnn/GoogLeNet/deploy.prototxt
    % url  = https://github.com/BVLC/caffe/raw/master/models/bvlc_googlenet/deploy.prototxt
    % hash = 7060345c8012294baa60eeb5901d2d3fd89d75fc
    %
    % ## Weights
    %
    % file = test/dnn/GoogLeNet/bvlc_googlenet.caffemodel
    % url  = http://dl.caffe.berkeleyvision.org/bvlc_googlenet.caffemodel
    % hash = 405fc5acd08a3bb12de8ee5e23a96bec22f08204
    % size = 51 MB
    %
    % ## Classes
    %
    % file = test/dnn/GoogLeNet/synset_words.txt
    % url  = https://github.com/opencv/opencv/raw/3.3.1/samples/data/dnn/synset_words.txt
    %

    dname = get_dnn_dir('GoogLeNet');
    net = cv.Net('Caffe', ...
        fullfile(dname, 'deploy.prototxt'), ...
        fullfile(dname, 'bvlc_googlenet.caffemodel'));
    labels = readLabels(fullfile(dname, 'synset_words.txt'), true);
    blobOpts = {'SwapRB',false, 'Size',[224 224], 'Mean',[104 117 123]};
end

function [net, labels, blobOpts] = ResNet()
    %RESNET  Deep Residual Networks, ResNet-50 [Caffe]
    %
    % homepage = https://github.com/KaimingHe/deep-residual-networks
    %
    % ## Model
    %
    % file = test/dnn/ResNet/ResNet-50-deploy.prototxt
    % url  = https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777
    % hash = 5d6fd5aeadd8d4684843c5028b4e5672b9e51638
    %
    % ## Weights
    %
    % file = test/dnn/ResNet/ResNet-50-model.caffemodel
    % url  = https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777
    % hash = b7c79ccc21ad0479cddc0dd78b1d20c4d722908d
    % size = 97.7 MB
    %
    % ## Classes
    %
    % file = test/dnn/ResNet/synset_words.txt
    % url  = https://github.com/opencv/opencv/raw/3.3.1/samples/data/dnn/synset_words.txt
    %

    dname = get_dnn_dir('ResNet');
    net = cv.Net('Caffe', ...
        fullfile(dname, 'ResNet-50-deploy.prototxt'), ...
        fullfile(dname, 'ResNet-50-model.caffemodel'));
    labels = readLabels(fullfile(dname, 'synset_words.txt'), true);
    blobOpts = {'SwapRB',false, 'Size',[224 224], 'Mean',[104 117 123]};
    %TODO: mean image from ResNet_mean.binaryproto
end

function [net, labels, blobOpts] = SqueezeNet(v)
    %SQUEEZENET  SqueezeNet [Caffe]
    %
    % homepage = https://github.com/DeepScale/SqueezeNet
    %
    % # SqueezeNet v1.0
    %
    % ## Model
    %
    % file = test/dnn/SqueezeNet/v1.0/deploy.prototxt
    % url  = https://github.com/DeepScale/SqueezeNet/raw/master/SqueezeNet_v1.0/deploy.prototxt
    % hash = 733249be856b9cd28ce929cd7c41874cf817c3c6
    %
    % ## Weights
    %
    % file = test/dnn/SqueezeNet/v1.0/squeezenet_v1.0.caffemodel
    % url  = https://github.com/DeepScale/SqueezeNet/raw/master/SqueezeNet_v1.0/squeezenet_v1.0.caffemodel
    % hash = 579d0beb658e43c45937bf8bb5e4034fea4e1f69
    % size = 4.76 MB
    %
    % # SqueezeNet v1.1
    %
    % ## Model
    %
    % file = test/dnn/SqueezeNet/v1.1/deploy.prototxt
    % url  = https://github.com/DeepScale/SqueezeNet/raw/master/SqueezeNet_v1.1/deploy.prototxt
    % hash = c226bbaaa4d83e2a3d3727618091e897e5f2e3aa
    %
    % ## Weights
    %
    % file = test/dnn/SqueezeNet/v1.1/squeezenet_v1.1.caffemodel
    % url  = https://github.com/DeepScale/SqueezeNet/raw/master/SqueezeNet_v1.1/squeezenet_v1.1.caffemodel
    % hash = 3397f026368a45ae236403ccc81cfcbe8ebe1bd0
    % size = 4.72 MB
    %
    % # Classes
    %
    % file = test/dnn/SqueezeNet/v1.x/synset_words.txt
    % url  = https://github.com/opencv/opencv/raw/3.3.1/samples/data/dnn/synset_words.txt
    %

    v = validatestring(v, {'v1.1', 'v1.0'});
    dname = get_dnn_dir(fullfile('SqueezeNet', v));
    if strcmp(v, 'v1.1')
        net = cv.Net('Caffe', ...
            fullfile(dname, 'deploy.prototxt'), ...
            fullfile(dname, 'squeezenet_v1.1.caffemodel'));
    else
        net = cv.Net('Caffe', ...
            fullfile(dname, 'deploy.prototxt'), ...
            fullfile(dname, 'squeezenet_v1.0.caffemodel'));
    end
    labels = readLabels(fullfile(dname, 'synset_words.txt'), true);
    blobOpts = {'SwapRB',false, 'Size',[227 227], 'Mean',[104 117 123]};
end

Load model...   Elapsed time is 0.197234 seconds.
Forward pass... Elapsed time is 0.113284 seconds.
Top-5 predictions:
 99.99% = space shuttle
  0.00% = airliner
  0.00% = submarine, pigboat, sub, U-boat
  0.00% = bullet train, bullet
  0.00% = missile