Method, electronic apparatus, and computer readable medium of constructing classifier for skin-infection detection

1. A method of constructing a classifier for melanoma detection comprising:

constructing a codebook of representative features based on a plurality of target-disease-irrelevant images;

extracting a plurality of transfer-learned target-disease features from a plurality of target-disease images according to the codebook, wherein the step of extracting the plurality of transfer-learned target-disease features comprises:

segmenting a target region from each of the target-disease images to correspondingly generate a segmented target-disease image; and

extracting the transfer-learned target-disease features from the segmented target-disease images; and

performing supervised learning based on the transfer-learned target-disease features to train the classifier for melanoma detection,

wherein the step of segmenting the target region from each of the target-disease images to correspondingly generate the segmented target-disease image comprises:

for each of the target-disease images:

initializing a contour at a center of the target region according to color;

evolving the contour such that a predefined energy function is minimized so as to obtain a terminated boundary; and

segmenting the target region from the target-disease image according to the terminated boundary to generate the segmented target-disease image.

2. A method of constructing a classifier for skin-infection detection comprising:

constructing a codebook of representative features based on a plurality of target-disease-irrelevant images;

extracting a plurality of transfer-learned target-disease features from a plurality of target-disease images according to the codebook, wherein the step of extracting the plurality of transfer-learned target-disease features comprises:

segmenting a target region from each of the target-disease images to correspondingly generate a segmented target-disease image; and

extracting the transfer-learned target-disease features from the segmented target-disease images; and

performing supervised learning based on the transfer-learned target-disease features to train the classifier for skin-infection detection,

wherein the step of segmenting the target region from each of the target-disease images to correspondingly generate the segmented target-disease image comprises:

for each of the target-disease images:

initializing a contour at a center of the target region according to color;

evolving the contour such that a predefined energy function is minimized so as to obtain a terminated boundary; and

segmenting the target region from the target-disease image according to the terminated boundary to generate the segmented target-disease image.

3. The method according to claim 2 , wherein the step of constructing the codebook of the representative features based on the target-disease-irrelevant images comprises:

constructing the codebook of the representative features based on the target-disease-irrelevant images by using a first representation learning model, wherein the first representation learning model comprises a plurality of layers of neurons, wherein the layers comprise at least one inner layer, and wherein activation values of the inner layers corresponds to the representative features of the target-disease-irrelevant images.

4. The method according to claim 3 , wherein the step of extracting the transfer-learned target-disease features from the target-disease images according to the codebook comprises:

for each of the target-disease images:

inputting the target-disease image into the first representation learning model;

propagating information in the target-disease image through the layers; and

obtaining the transfer-learned features of the target-disease image according to at least one of the inner layers.

5. The method according to claim 2 , wherein the step of constructing the codebook of the representative features based on the target-disease-irrelevant images comprises:

constructing the codebook of the representative features based on the target-disease-irrelevant images by using a second representation learning model, wherein the second representation learning model comprises a plurality of layers of units, wherein the layers comprise at least one hidden layer having a plurality of hidden units, and wherein probabilistic values of the hidden units correspond to the representative features of the target-disease-irrelevant images.

6. The method according to claim 5 , wherein the step of extracting the transfer-learned features from the target-disease images according to the codebook comprises:

for each of the target-disease images:

inputting the target-disease image into the second representation learning model;

propagating information in the target-disease image through the layers; and

obtaining the transfer-learned features of the target-disease image according to at least one of the hidden units.

7. The method according to claim 2 , wherein after the step of performing supervised learning on at least the transfer-learned features to train the classifier for skin-infection detection, the method further comprises:

capturing a new image by an image capturing device; and

classifying whether the new image corresponds to a skin infection condition by using the classifier for skin-infection detection to generate a classification result; and

outputting the classification result.

8. An electronic apparatus comprising:

a communication interface;

a memory, recording a plurality of modules; and

one or a plurality of processors, coupled to the communication interface and the memory, and accessing and executing the modules stored in the memory, wherein the modules comprise:

a codebook constructing module, constructing a codebook of representative features based on a plurality of target-disease-irrelevant images obtained via the communication interface;

a feature extracting module, extracting a plurality of transfer-learned target-disease features from a plurality of target-disease images obtained via the communication interface according to the codebook, wherein the feature extracting module segments a target region from each of the target-disease images to correspondingly generate a segmented target-disease image, and extracts the transfer-learned target-disease features from the segmented target-disease images; and

a classifier training module, performing supervised learning based on the transfer-learned target-disease features to train the classifier for skin-infection detection,

wherein for each of the target-disease images, the feature extracting module initializes a contour at a center of the target region according to color, evolves the contour such that the energy function is minimized so as to obtain a terminated boundary, and segments the target region from the target-disease image according to the terminated boundary to generate the segmented target-disease image.

9. The electronic apparatus according to claim 8 , wherein the codebook constructing module constructs the codebook based on the target-disease-irrelevant images by using a first representation learning model, wherein the first representation learning model comprises a plurality of layers of neurons, wherein the layers comprise at least one inner layer, and wherein activation values of the inner layers corresponds to the representative features of the target-disease-irrelevant images.

10. The electronic apparatus according to claim 9 , wherein for each of the target-disease images, the feature extracting module inputs the target-disease image into the first representation learning model, propagates information in the target-disease image through the layers, and obtains the transfer-learned features of the target-disease image according to at least one of the inner layers.

11. The electronic apparatus according to claim 9 , wherein the classifier obtains a new image captured by an image capturing device via the communication interface, classifies whether the new image corresponds to a skin infection condition to generate a classification result, and outputs the classification result.

12. The electronic apparatus according to claim 8 , wherein the codebook constructing module constructs the codebook based on the target-disease-irrelevant images by using a second representation learning model, wherein the second representation learning model comprises a plurality of layers of units, wherein the layers comprise at least one hidden layer having a plurality of hidden units, and wherein probabilistic values of the hidden units correspond to the representative features of the target-disease-irrelevant images.

13. The electronic apparatus according to claim 12 , wherein for each of the target-disease images, the feature extracting module inputs the target-disease image into the second representation learning model, propagates information in the target-disease image through the layers, and obtains the transfer-learned features of the target-disease image according to at least one of the hidden units.

14. A non-transitory computer readable medium, storing programs to be loaded into an electronic apparatus to perform steps of:

constructing a codebook of representative features based on a plurality of target-disease-irrelevant images;

extracting a plurality of transfer-learned target-disease features from a plurality of target-disease images according to the codebook, wherein the step of extracting the plurality of transfer-learned target-disease features comprises:

segmenting a target region from each of the target-disease images to correspondingly generate a segmented target-disease image; and

extracting the transfer-learned target-disease features from the segmented target-disease images;

performing supervised learning based on the transfer-learned target-disease features to train the classifier for skin-infection detection,

wherein the step of segmenting the target region from each of the target-disease images to correspondingly generate the segmented target-disease image comprises:

for each of the target-disease images:

initializing a contour at a center of the target region according to color;

evolving the contour such that a predefined energy function is minimized so as to obtain a terminated boundary; and

segmenting the target region from the target-disease image according to the terminated boundary to generate the segmented target-disease image.