Computer-implemented method of enhancing object detection in a digital image of known underlying structure, and corresponding module, data processing apparatus and computer program

1 . A computer-implemented method of enhancing object detection in a digital image of known underlying structure using pre-processed images with underlying structure and with any objects detected and bounding boxes inserted over the objects, the method comprising:

extracting or generating images with the underlying structure but without objects detected as negative images;

extracting images with the underlying structure and with an object detected as positive images;

inputting pairs of negative and positive images through a feature extraction section in a neural network to extract feature vectors of the images;

contrasting feature vectors of each pair of positive and negative images and thereby provide a contrast vector and gating the result to form an attention vector; and

processing the attention vector and the feature vector of the positive image to produce an output.

2 . The method according to claim 1 , further comprising: before extracting the images, dividing original images into tiles, each tile forming a potential image with underlying structure for input.

3 . The method according to claim 2 , wherein the tiles are pre-processed to align the underling structure.

4 . The method according to claim 1 , wherein gating the subtracted feature vectors of the pair of positive and negative images comprises:

using a gating block to provide a learnable block to optimise the attention vector and reduce the noise from potential misalignment of the underlying structure in the positive and negative image.

5 . The method according to claim 4 , wherein the gating block is a convolution block with a sigmoid activation function.

6 . The method according to claim 1 , wherein processing the attention vector and the feature vector of the positive image comprises:

multiplying the attention vector and the feature vector of the positive image to provide a gain vector and summing the gain vector and the feature vector of the positive image to produce the output as an enhanced feature vector.

7 . The method according to claim 1 , wherein the output is fed to a pooling layer of the neural network.

8 . The method according to claim 1 , wherein images are input from an encoder layer into the method.

9 . The method according to claim 1 , wherein the underlying structure is a medical topological structure of a medical image of a body part.

10 . The method according to claim 1 , wherein the underlying structure contains consistencies in spatial relations between objects in an image.

11 . The method according to claim 10 , wherein the wherein the objects in the image are objects within an organ or body parts.

12 . The method according to claim 9 , wherein the images are of a retina, and the images are pre-processed to flatten the curve of the retina.

13 . The method according to claim 1 , carried out in a YOLO network, which learns the bounding box and classification of the object in the bounding box in a single stage.

14 . The method according to claim 1 , carried out in a training phase of neural network, in which each image is weighted according to the number of classes of object included, and the weighting is used to select images for use in training.

15 . A module in a neural network which is configured to carry out a method of enhancing object detection in a digital image of known underlying structure using pre-processed images with underlying structure and with any objects detected and bounding boxes inserted over the objects, the method comprising:

extracting or generating images with the underlying structure but without objects detected as negative images;

extracting images with the underlying structure and with an object detected as positive images;

inputting pairs of negative and positive images through a feature extraction section in a neural network to extract feature vectors of the images;

contrasting feature vectors of each pair of positive and negative images and thereby provide a contrast vector and gating the result to form an attention vector; and

processing the attention vector and the feature vector of the positive image to produce an output.

16 . The module of claim 15 , wherein the module is positioned in the neural network after an encoder module and before a pooling module.

17 . A data processing apparatus comprising: a memory; and a processor, wherein the memory comprises instructions for enhancing object detection in a digital image of known underlying structure using pre-processed images with underlying structure and with any objects detected and bounding boxes inserted over the objects, and

the processor is configured to execute the instructions to:

extract or generate images with the underlying structure but without objects detected as negative images;

extract images with the underlying structure and with an object detected as positive images;

input pairs of negative and positive images through a feature extraction section in a neural network to extract feature vectors of the images;

contrast feature vectors of each pair of positive and negative images and thereby provide a contrast vector and gating the result to form an attention vector; and

process the attention vector and the feature vector of the positive image to produce an output.