DETECTING INFECTION OF PLANT DISEASES WITH IMPROVED MACHINE LEARNING

• US 20200134358A1
• Filed: 10/24/2019
• Published: 04/30/2020
• Est. Priority Date: 10/24/2018
• Status: Active Grant

• Current Assignee: Climate Pty., Ltd.
• Original Assignee: The Climate Corporation (Bayer AG)
• Inventors: SHE, YING, GUAN, WEI
• Granted Patent: US 10,956,780 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 18/2113 by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation
• CPC Class Codes: G06F 18/217 Validation; Performance evaluation; Active pattern learning techniques
• CPC Class Codes: G06F 18/2431 Multiple classes
• CPC Class Codes: G06N 20/20 Ensemble learning
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/086 using evolutionary algorithms, e.g. genetic algorithms or genetic programming
• CPC Class Codes: G06N 3/126 Evolutionary algorithms, e.g. genetic algorithms or genetic programming
• CPC Class Codes: G06N 5/01 Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
• CPC Class Codes: G06V 10/771 Feature selection, e.g. selecting representative features from a multi-dimensional feature space
• CPC Class Codes: G06V 10/776 Validation; Performance evaluation
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 20/60 Type of objects
• CPC Class Codes: G06V 20/68 Food, e.g. fruit or vegetables

Abstract

A system and processing methods for refining a convolutional neural network (CNN) to capture characterizing features of different classes are disclosed. In some embodiments, the system is programmed to start with the filters in one of the last few convolutional layers of the initial CNN, which often correspond to more class-specific features, rank them to hone in on more relevant filters, and update the initial CNN by turning off the less relevant filters in that one convolutional layer. The result is often a more generalized CNN that is rid of certain filters that do not help characterize the classes.

Claims

1. A computer-implemented method of refining a convolutional neural network to focus on class-specific features, comprising:

receiving, by a processor, digital data representing an initial convolutional neural network (CNN) comprising a series of convolution blocks,

each of the series of convolution blocks comprising a convolutional layer having one or more filters,

a convolutional layer of a distinct number of last convolution blocks of the series of convolution blocks having a certain number of filters corresponding to a certain number of features;

receiving, by the processor, a set of digital images and a corresponding set of class labels each identifying a class of a plurality of classes;

processing each digital image in the set of digital images using the series of convolution blocks to generate a specific number of feature maps for each of the set of digital images;

generating a vector for each of the set of digital images based on the specific number of features maps for the digital image;

ranking the certain number of filters based on the set of vectors for the set of digital images and the corresponding set of class labels;

selecting a particular number of highest-ranking filters from the certain number of filters;

constructing an updated CNN from the initial CNN to eliminate application of non-selected filters of the certain number of filters;

applying the updated CNN to a new image received from a client device to obtain a classification of the new image into one of the plurality of classes;

transmitting information related to the classification.

2. The computer-implemented method of claim 1,

one of the set of digital images is a photo of a type of plant,

one of the plurality of classes is a disease likely to affect the type of plant.

3. The computer-implemented method of claim 1, each of the series of convolution blocks further comprising a pooling layer.

4. The computer-implemented method of claim 1, the certain number being larger than a number of filters associated with a convolutional layer of one of the series of convolution blocks other than the distinct number of last convolution blocks.

5. The computer-implemented method of claim 1, the generating comprising computing a component of the vector as an aggregate over values in a feature map of the specific number of feature maps for the digital image.

6. The computer-implemented method of claim 1,

the specific number of feature maps being equal to the certain number of filters,

the generating comprising performing global average pooling on each of the specific number of feature maps for the digital image.

7. The computer-implemented method of claim 1, the selecting being based on past results associated with different number of highest-ranking filters selected for constructing an updated CNN.

8. The computer-implemented method of claim 1, the ranking comprising building a random forest from the set of vectors and the associated set of class labels.

9. The computer-implemented method of claim 1, the updated CNN comprising the series of convolution blocks and a masking layer that masks out non-selected filters of the certain number of filters.

10. The computer-implemented method of claim 1, the constructing comprising modifying the convolutional layer to remove non-selected filters from the certain number of filters.

11. The computer-implemented method of claim 1, the initial CNN further comprising a classification block after the series of convolution blocks, the classification block including a fully-connected layer.

12. The computer-implemented method of claim 11, the updated CNN comprising an updated classification block, the updated classification block including the fully-connected layer retaining weights associated with the particular number of highest-ranking filters.

13. The computer-implemented method of claim 12, the updated classification block further including a softmax layer.

16. The computer-implemented method of claim 1, the selecting being based on past results associated with different number of highest-ranking filters selected for constructing an updated CNN.

14. One or more non-transitory computer-readable media storing one or more sequences of instructions which when executed using one or more processors cause the one or more processors to execute a method of refining a convolutional neural network to focus on class-specific features, the method comprising:

receiving digital data representing an initial convolutional neural network (CNN) comprising a series of convolution blocks,

each of the series of convolution blocks comprising a convolutional layer having one or more filters,

a convolutional layer of a distinct number of last convolution blocks of the series of convolution blocks having a certain number of filters corresponding to a certain number of features;

receiving a set of digital images and a corresponding set of class labels each identifying a class of a plurality of classes;

processing each digital image in the set of digital images using the series of convolution blocks to generate a specific number of feature maps for each of the set of digital images;

generating a vector for each of the set of digital images based on the specific number of features maps for the digital image;

ranking the certain number of filters based on the set of vectors for the set of digital images and the corresponding set of class labels;

selecting a particular number of highest-ranking filters from the certain number of filters;

constructing an updated CNN from the initial CNN to eliminate application of non-selected filters of the certain number of filters;

applying the updated CNN to a new image received from a client device to obtain a classification of the new image into one of the plurality of classes;

transmitting information related to the classification.

15. The one or more non-transitory computer-readable media of claim 14, the generating comprising computing a component of the vector as an aggregate over values in a feature map of the specific number of feature maps for the digital image.

17. The one or more non-transitory computer-readable media of claim 14, the ranking comprising building a random forest from the set of vectors and the associated set of class labels.

18. The one or more non-transitory computer-readable media of claim 14, the updated CNN comprising the series of convolution blocks and a masking layer that masks out non-selected filters of the certain number of filters.

19. The one or more non-transitory computer-readable media of claim 14, the initial CNN further comprising a classification block after the series of convolution blocks, the classification block including a full-connected layer.

20. The one or more non-transitory computer-readable media of claim 19, the updated CNN comprising an updated classification block, the updated classification block including the fully-connected layer retaining weights associated with the particular number of highest-ranking filters.