CN-121982437-A - Feature extraction method and system based on 2.5D convolution

CN121982437ACN 121982437 ACN121982437 ACN 121982437ACN-121982437-A

Abstract

The invention discloses a feature extraction method and a feature extraction system based on 2.5D convolution, which are used for cutting useless areas of an original 3D image to obtain a preprocessed 3D image, determining a plurality of slicing directions according to the object outline of the original 3D image, dividing the preprocessed 3D image into a plurality of 2D image sets according to the plurality of slicing directions, carrying out fusion processing on all 2D images in the plurality of 2D image sets to obtain a 2.5D image output set, carrying out multichannel processing of a convolution neural network on the 2.5D image output set to extract internal and external structural features of the object, and determining abnormal areas of the object according to the structural features. By carrying out clipping pretreatment and multi-direction slicing segmentation on the 3D image, a 2.5D image set is obtained, non-time dimension related processing on the 3D image is realized, the computational complexity and the computational effort are reduced, the memory resource requirement is met, and the image recognition efficiency and the reliability are improved.

Inventors

YU DAN
WANG DANXING
Xing Zhihuan

Assignees

慧之安信息技术股份有限公司

Dates

Publication Date: 20260505
Application Date: 20251212

Claims (8)

1. A feature extraction method based on 2.5D convolution, comprising: cutting useless areas of an original 3D image to obtain a preprocessed 3D image, and determining a plurality of slicing directions according to the object contour of the original 3D image; dividing the preprocessing 3D image into a plurality of 2D image sets according to the slicing directions, and carrying out fusion processing on all 2D images in the plurality of 2D image sets to obtain a 2.5D image output set; and carrying out multichannel processing of a convolutional neural network on the 2.5D image output set, extracting to obtain the internal and external structural features of the object, and determining the abnormal region of the object according to the structural features.
2. The feature extraction method based on 2.5D convolution as claimed in claim 1, wherein: And determining a plurality of slicing directions according to the object contour of the original 3D image, wherein the steps comprise: Carrying out multichannel information distribution identification on an original 3D image to obtain an information distribution state of the original 3D image, wherein the information distribution state refers to a pixel chromaticity information amount airspace distribution state and a pixel texture information amount airspace distribution state; determining an useless area of the original 3D image according to the information distribution state, and cutting the original 3D image into a preprocessed 3D image with a regular boundary according to the boundary of the useless area; Extracting object contour data from the original 3D image, and determining the spatial distribution of the object contour information quantity of the original 3D image according to the object contour data so as to determine a plurality of slicing directions, wherein the slicing directions refer to directions in which the spatial density of the object contour information quantity in the original 3D image exceeds a preset density.
3. The feature extraction method based on 2.5D convolution as claimed in claim 1, wherein: Dividing the preprocessing 3D image into a plurality of 2D image sets according to the plurality of slicing directions, and carrying out fusion processing on all 2D images in the plurality of 2D image sets to obtain a 2.5D image output set, wherein the method comprises the following steps: Uniformly dividing the preprocessed 3D image into a plurality of 2D images along each slice direction, and carrying out noise reduction filtering preprocessing and sequential arrangement preprocessing on the plurality of 2D images to obtain a 2D image set, wherein the preprocessed 3D images are divided along each slice direction and then are in one-to-one correspondence to obtain the corresponding 2D image set; and according to the slicing directions corresponding to all the 2D image sets, carrying out volume fusion processing and triple-focus cross-dimension interaction mechanism processing on all the 2.5D image sets to obtain a 2.5D image output set.
4. The feature extraction method based on 2.5D convolution as claimed in claim 1, wherein: The method comprises the steps of carrying out multichannel processing of a convolutional neural network on the 2.5D image output set, extracting to obtain structural features of the inside and the outside of an object, and determining an abnormal region of the object according to the structural features, wherein the method comprises the following steps: Capturing the attention weights of the channel, the height, the channel, the width and the space dimension of the 2.5D image set through a plurality of branches subordinate to a convolutional neural network, demarcating a key region of the 2.5D image output set according to the attention weights, and extracting from the key region to obtain the internal and external structural characteristics of an object; And determining an abnormal region of the object according to the construction deviation and the distribution position thereof, wherein the abnormal region refers to a structure missing region existing inside or outside the object.
5. A feature extraction system based on 2.5D convolution, comprising: The image clipping module is used for clipping useless areas of the original 3D image to obtain a preprocessed 3D image; the slice direction determining module is used for determining a plurality of slice directions according to the object contour of the original 3D image; the segmentation module is used for respectively segmenting the preprocessing 3D image into a plurality of 2D image sets according to the plurality of slicing directions; The fusion module is used for carrying out fusion processing on all 2D images in the plurality of 2D image sets to obtain a 2.5D image output set; The convolution processing module is used for carrying out multichannel processing of a convolution neural network on the 2.5D image output set and extracting to obtain the internal and external structural features of the object; and the abnormal region determining module is used for determining the abnormal region of the object according to the construction characteristics.
6. The 2.5D convolution-based feature extraction system according to claim 5, wherein: The image clipping module is used for clipping useless areas of the original 3D image to obtain a preprocessed 3D image, and comprises the following steps: Carrying out multichannel information distribution identification on an original 3D image to obtain an information distribution state of the original 3D image, wherein the information distribution state refers to a pixel chromaticity information amount airspace distribution state and a pixel texture information amount airspace distribution state; determining an useless area of the original 3D image according to the information distribution state, and cutting the original 3D image into a preprocessed 3D image with a regular boundary according to the boundary of the useless area; The slice direction determining module is configured to determine a plurality of slice directions according to an object contour of the original 3D image, including: Extracting object contour data from the original 3D image, and determining the spatial distribution of the object contour information quantity of the original 3D image according to the object contour data so as to determine a plurality of slicing directions, wherein the slicing directions refer to directions in which the spatial density of the object contour information quantity in the original 3D image exceeds a preset density.
7. The 2.5D convolution-based feature extraction system according to claim 5, wherein: The segmentation module is used for respectively segmenting the preprocessing 3D image into a plurality of 2D image sets according to the plurality of slicing directions, and comprises the following steps: Uniformly dividing the preprocessed 3D image into a plurality of 2D images along each slice direction, and carrying out noise reduction filtering preprocessing and sequential arrangement preprocessing on the plurality of 2D images to obtain a 2D image set, wherein the preprocessed 3D images are divided along each slice direction and then are in one-to-one correspondence to obtain the corresponding 2D image set; The fusion module is used for carrying out fusion processing on all 2D images in the plurality of 2D image sets to obtain a 2.5D image output set, and comprises the following steps: and according to the slicing directions corresponding to all the 2D image sets, carrying out volume fusion processing and triple-focus cross-dimension interaction mechanism processing on all the 2.5D image sets to obtain a 2.5D image output set.
8. The 2.5D convolution-based feature extraction system according to claim 5, wherein: The convolution processing module is used for performing multichannel processing of a convolution neural network on the 2.5D image output set, extracting and obtaining structural features of the interior and the exterior of an object, and comprises the following steps: Capturing the attention weights of the channel, the height, the channel, the width and the space dimension of the 2.5D image set through a plurality of branches subordinate to a convolutional neural network, demarcating a key region of the 2.5D image output set according to the attention weights, and extracting from the key region to obtain the internal and external structural characteristics of an object; the abnormal region determining module is configured to determine an abnormal region of an object according to the structural feature, and includes: And determining an abnormal region of the object according to the construction deviation and the distribution position thereof, wherein the abnormal region refers to a structure missing region existing inside or outside the object.

Description

Feature extraction method and system based on 2.5D convolution Technical Field The invention relates to the technical field of image processing, in particular to a feature extraction method and system based on 2.5D convolution. Background The 3D convolution is a model that extends the traditional 2D convolution operation to the temporal dimension, which is capable of directly manipulating a sequence of video frames, extracting spatio-temporal features from video data and capturing motion information by the 3D convolution layer while considering both temporal and spatial dimension features. The 3D convolution is mainly applied to dynamic video recognition, which requires greater computational resources and memory space during training and reasoning. In the actual image data processing, not all scenes relate to dynamic video, but only static three-dimensional images, and if the 3D convolution processing is used, calculation force and storage resources are wasted, and if the 2D convolution processing is used, accurate static three-dimensional image recognition cannot be realized. Therefore, how to accurately identify the static three-dimensional image under the condition of efficiently utilizing computing power and memory resources has great significance for improving the image identification efficiency and the credibility. Disclosure of Invention Considering that the existing 3D convolution can cause a great deal of computational effort and idle waste of memory resources when identifying static three-dimensional images, the adoption of 2D convolution to identify the static three-dimensional images can not accurately determine the three-dimensional details in the static three-dimensional images, and the image identification efficiency and the reliability are reduced. The present invention has been made in view of the above problems, and has as its object to provide a 2.5D convolution-based feature extraction method that overcomes or at least partially solves the above problems, including: cutting useless areas of an original 3D image to obtain a preprocessed 3D image, and determining a plurality of slicing directions according to the object contour of the original 3D image; dividing the preprocessing 3D image into a plurality of 2D image sets according to the slicing directions, and carrying out fusion processing on all 2D images in the plurality of 2D image sets to obtain a 2.5D image output set; and carrying out multichannel processing of a convolutional neural network on the 2.5D image output set, extracting to obtain the internal and external structural features of the object, and determining the abnormal region of the object according to the structural features. Optionally, the method comprises the steps of carrying out useless region clipping on an original 3D image to obtain a preprocessed 3D image, determining a plurality of slicing directions according to the object contour of the original 3D image, wherein the steps comprise: Carrying out multichannel information distribution identification on an original 3D image to obtain an information distribution state of the original 3D image, wherein the information distribution state refers to a pixel chromaticity information amount airspace distribution state and a pixel texture information amount airspace distribution state; determining an useless area of the original 3D image according to the information distribution state, and cutting the original 3D image into a preprocessed 3D image with a regular boundary according to the boundary of the useless area; Extracting object contour data from the original 3D image, and determining the spatial distribution of the object contour information quantity of the original 3D image according to the object contour data so as to determine a plurality of slicing directions, wherein the slicing directions refer to directions in which the spatial density of the object contour information quantity in the original 3D image exceeds a preset density. Optionally, according to the slicing directions, the preprocessing 3D image is respectively divided into a plurality of 2D image sets, and fusion processing is carried out on all 2D images in the 2D image sets to obtain a 2.5D image output set, which comprises the following steps: Uniformly dividing the preprocessed 3D image into a plurality of 2D images along each slice direction, and carrying out noise reduction filtering preprocessing and sequential arrangement preprocessing on the plurality of 2D images to obtain a 2D image set, wherein the preprocessed 3D images are divided along each slice direction and then are in one-to-one correspondence to obtain the corresponding 2D image set; and according to the slicing directions corresponding to all the 2D image sets, carrying out volume fusion processing and triple-focus cross-dimension interaction mechanism processing on all the 2.5D image sets to obtain a 2.5D image output set. Optionally, carrying out multichannel processing of