Explainable AI with Score-CAM

버전 1.1.0 (13.3 MB) 작성자: Kenta
This repo shows a MATLAB implementation of score-CAM (Wang et al., CVPR workshop, 2020). This method is the first gradient-free CAM-based vi
다운로드 수: 196
업데이트 날짜: 2021/12/16

View Explainable AI with Score-CAM on File Exchange

Explainable AI: interpreting the classification using score-CAM


This example shows how to use score-CAM (class activation mapping) [1] to investigate and explain the predictions of a deep convolutional neural network for image classification. The algorithm like CAM [2] and grad-CAM [3] are popular for explainable AI with image classification, but Score-CAM proposed a "gradient-free" method for the explanation. The original implementation can be found here.




Keywords: classification, deep learning, explainable AI, gradient, Score-CAM

[1] Wang, H., Du, M., Yang, F. and Zhang, Z., 2019. Score-CAM: Improved Visual Explanations Via Score-Weighted Class Activation Mapping. arXiv preprint arXiv:1910.01279.

[2] Zhou, B., Khosla, A., Lapedriza, A., Oliva, A. and Torralba, A., 2016. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2921-2929).

[3] Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D. and Batra, D., 2017. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (pp. 618-626).

Load Pretrained Network

Load a pretrained convolutional neural network for image classification. Please select the pretrained network from the pull-down. In this code, the network such as SqueezeNet, GoogLeNet, ResNet-18, and MobileNet-v2 are avalable. While we cannot perform class activation mapping with networks that have multiple fully connected layers at the end of the network, such as AlexNet, VGG-16, and VGG-19, score-CAM can visualize the region where is influencial to the prediction .

clear;clc;close all
netName = "squeezenet";
net = eval(netName);

Extract the image input size and the output classes of the network

The activationLayerName helper function, defined at the end of this example, returns the name of the layer to extract the activations from. This layer is the ReLU layer that follows the last convolutional layer of the network.

inputSize = net.Layers(1).InputSize(1:2);
classes = net.Layers(end).Classes;
layerName = activationLayerName(netName);

Read your target image and resize so as to feed into the CNN network

Compute the activations of the resized image in the ReLU layer that follows the last convolutional layer of the network.

im = imread('CatImg.png');
imResized = imresize(im,[inputSize(1:2)]);

Classify the target image into a category

PredCategory = 
     Norwegian elkhound 

Extract the index of the predicted class for the following process

classIdx=find((net.Layers(end, 1).Classes==PredCategory)==1);

Calculate feature map using activation function

Display the feature maps after performing convolution to the final convolutional layer

featureMap = activations(net,imResized,layerName);
idx=randi(MapDim,[1 1]);
figure;imagesc(normalize(featureMap(:,:,idx),'range'));title('example of the feature map');colorbar


Obtain dot-product of input image and resized feature map

This section includes

  1. normalize the feature maps into [0 1]
  2. convert the feature map into the map with three-channel (grayscale to RGB imge).
  3. Obtain dot-product of the input image and the feature map in the last step
for i=1:size(featureMap,3)
    featureMap_normalized(:,:,i) = normalize(featureMap_resized(:,:,i),'range');

Display the example

figure;idx=randi(MapDim,[12 1]);
montage(uint8(exImg));title('example of masked images for input')


Inference with the masked images

Use classify function to predict and extract the score for the target class

% specify minibatch size. Return an error if the memory is not enough   
% score: (the number of test image)-by-(the number of class (1000))

Multiply the featuremap with the score with corresponding input mask image

Generate course heat map for your final result. The size of the course map is identical to the size of the feature map.

score_CAM_prep=featureMap_normalized.*reshape(CIC_norm,[1 1 numel(CIC_norm)]);

Operate ReLu function where the element less than 0 is converted into 0. Only the region which is positively influencial to the class of interest is extracted for the linear map combination.


Normalize the feature map into [0 1]. Just writing 'range' is fine for the [0 1] normalization.


Visualization setup

cmap = jet(255).*linspace(0,1,255)';
score_CAM = ind2rgb(uint8(score_CAM_scaled*255),cmap)*255;

Combine the grayscaled input image with the heat map

combinedImage = double(rgb2gray(imResized))/2 + score_CAM;

Normalize into [0 255] for image output

combinedImage = normalizeImage(combinedImage)*255;

Display the score-cam result



Helper Functions

function N = normalizeImage(I)
minimum = min(I(:));
maximum = max(I(:));
N = (I-minimum)/(maximum-minimum);

Get the name of the last convolutional layer

Input: the pretrained network name

Output: the name of the last convolutional layer

function layerName = activationLayerName(netName)

if netName == "squeezenet"
    layerName = 'conv10';
elseif netName == "googlenet"
    layerName = 'inception_5b-output';
elseif netName == "resnet18"
    layerName = 'res5b';
elseif netName == "mobilenetv2"
    layerName = 'Conv_1';
elseif netName == "vgg16"
    layerName = 'conv5_3';
elseif netName == "alexnet"
    layerName = 'conv5';


인용 양식

Kenta (2024). Explainable AI with Score-CAM (https://github.com/KentaItakura/Explainable-AI-interpreting-the-classification-using-score-CAM/releases/tag/1.1.0), GitHub. 검색됨 .

MATLAB 릴리스 호환 정보
개발 환경: R2021b
모든 릴리스와 호환
플랫폼 호환성
Windows macOS Linux

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!
버전 게시됨 릴리스 정보

이 GitHub 애드온의 문제를 보거나 보고하려면 GitHub 리포지토리로 가십시오.
이 GitHub 애드온의 문제를 보거나 보고하려면 GitHub 리포지토리로 가십시오.