A system and method suitable for detecting changes in a building

Abstract

The application discloses a system and method suitable for detecting building changes, and belongs to the technical field of information processing. The system comprises a data acquisition module, the data acquisition module is used for acquiring a high-resolution remote sensing image of a target building, and taking the high-resolution remote sensing image as a training set of an input data enhancement module; a data enhancement module, the data enhancement module trains the training set, acquires a to-be-processed image, and inputs the to-be-processed image into a convolutional neural network, pre-processes the to-be-processed image through the convolutional neural network, and acquires a feature decoding graph of the to-be-processed image; and a detection module, the detection module determines a change condition of the building according to the feature decoding graph. Through the processing of the data enhancement module on the image, the system can effectively avoid the problem of poor fitting effect of a training model caused by too many negative samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and more particularly, to a system and method suitable for detecting building changes. BACKGROUND

[0002] The target of high-resolution remote sensing building change detection task is to predict whether there is a new building or a building removal at the pixel level.

[0003] The existing change detection system can be divided into two categories: traditional manual feature-based and deep feature-based methods. The traditional method mainly represents the features according to human settings, and distinguishes them according to the pixel points and their domain pixel points. In the past period of time, these methods have achieved good results. However, with the rise of deep learning, the disadvantages of traditional methods relying on human factors have been shown, while deep learning can objectively and autonomously learn beneficial features, greatly improving the F1 score of change detection.

[0004] The deep feature-based method adopts a deep learning change detection framework, which generally includes the following aspects: a data enhancement module (to solve the imbalance of positive and negative samples in remote sensing images), an image convolution feature extraction (encoder), a feature fusion (to fuse features of different time phases), and a feature decoding (pixel point category prediction). Among them, the feature fusion mainly exists in two forms: 1) the preprocessed different time phase images are spliced before being input into the network; 2) the features of different time phase images are extracted respectively, and then the obtained features are spliced. The decoder and the encoder part transmit information through a jump connection. The existing change detection modules will be described in detail below.

[0005] The first is IAug_CDNet, proposed by Chen et al. in "H.Chen, W.Li, and Z.Shi, Adversarialinstance augmentation for building change detection in remote sensing images, IEEE Transactions on Geoscience and Remote Sensing, DOI:10.1109 / TGRS.2021.3066802,2021." This work proposes a novel data-level solution to address the problem of insufficient training data due to sample sparsity. Called Instance-Level Change Augmentation (IAug), it generates bitemporal images containing a large number of different building changes through generative adversarial networks (GANs). The key to IAug is blending synthetic building instances into appropriate locations within a bitemporal image, further proposing a context-aware, realistic blending of buildings and background. This method utilizes GANs to construct similar sample buildings, increasing the number of positive samples in the dataset. However, when blending synthetic building instances into bitemporal images, it is difficult to accurately select appropriate locations, resulting in inconsistencies between buildings and background.

[0006] The second is DCA-Det, proposed by Zhang et al. in "L. Zhang, X. Hu, M. Zhang, Z. Shu and H. Zhou, Object-level change detection with a dual correlation attention-guided detector, ISPRS Journal of Photogrammetry and Remote Sensing, vol. 177, pp. 147-160, 2021." This work proposes an alternating mosaic data augmentation method that randomly selects four images from the dataset, merges them to generate a new image pair for data augmentation. However, when the changing building is located at the edge of the image merge, it cannot effectively determine the change based on contextual information. Summary of the Invention

[0007] To address the above problems, this invention proposes a system suitable for detecting changes in buildings, comprising:

[0008] A data acquisition module is used to acquire high-resolution remote sensing images of the target building and use the high-resolution remote sensing images as the training set for the input data augmentation module.

[0009] The data enhancement module trains a training set, obtains a to-be-processed image, inputs the to-be-processed image into a convolutional neural network, pre-processes the to-be-processed image through the convolutional neural network, and obtains a feature decoding graph of the to-be-processed image.

[0010] The detection module determines the change of the building according to the feature decoding graph.

[0011] Optionally, the data enhancement module trains a training set, obtains a to-be-processed image, and includes the following steps.

[0012] The high-resolution remote sensing image of the target building is partitioned according to a preset size, and a plurality of sub-regions are obtained.

[0013] An arbitrary point is randomly selected in the sub-region, and the sub-region is cropped with the arbitrary point as the center, and the cropped image is taken as a candidate image.

[0014] Different probability weights are given to the candidate images according to the number of positive samples of the candidate images, so as to determine the probability of being selected by the candidate images, and the candidate images with a large probability of being selected are taken as to-be-processed images.

[0015] Optionally, the convolutional neural network includes a convolutional feature encoding layer, a multi-stage information fusion layer and a feature decoding layer.

[0016] The convolutional feature encoding layer uses a large convolution kernel to extract image shallow features of the to-be-processed image, obtains time phase 1 shallow features and time phase 2 shallow features, and splices the time phase 1 shallow features and the time phase 2 shallow features to obtain a preliminary fusion feature. The time phase 1 shallow features, the time phase 2 shallow features and the preliminary fusion feature are respectively input into a twin network, and a time phase 1 feature map, a time phase 2 feature map and a preliminary fusion feature map of the to-be-processed image are obtained through multi-layer convolution.

[0017] The multi-stage information fusion layer splices the time phase 1 feature map and the time phase 2 feature map at the same resolution after difference, addition and splicing, and splices the preliminary fusion feature map. After splicing is completed, channel attention is used for information screening in the channel dimension to obtain a medium-term fusion feature map.

[0018] The feature decoding layer decodes the time phase 1 feature map, the time phase 2 feature map and the medium-term fusion feature map to obtain a time phase 1 decoding feature map, a time phase 2 decoding feature map and a medium-term fusion decoding feature map. The time phase 1 decoding feature map and the time phase 2 decoding feature map are taken as supplementary information, the supplementary information is fused with the medium-term fusion decoding feature map, and a feature decoding graph of the to-be-processed image is obtained.

[0019] Optionally, the convolutional neural network is an AlexNet network, a GoogLeNet network, a VGG network or a ResNe network.

[0020] The application further provides a method for detecting changes of a building, comprising:

[0021] collecting a high-resolution remote sensing image of a target building, and taking the high-resolution remote sensing image as a training set;

[0022] training the training set, obtaining a to-be-processed image, and inputting the to-be-processed image into a convolutional neural network, pre-processing the to-be-processed image through the convolutional neural network, and obtaining a feature decoding image of the to-be-processed image;

[0023] determining a change of the building according to the feature decoding image.

[0024] Optionally, the training of the training set and the obtaining of the to-be-processed image comprise:

[0025] partitioning the high-resolution remote sensing image of the target building according to a preset size, and obtaining a plurality of sub-regions;

[0026] randomly selecting an arbitrary point in the sub-region, and cutting the sub-region with the arbitrary point as a center, taking the cut image as a candidate image;

[0027] assigning different probability weights to the candidate images according to the number of positive samples of the candidate images, to determine the probability of being selected by the candidate images, and taking the candidate images with a large probability of being selected as the to-be-processed images.

[0028] Optionally, the convolutional neural network comprises a convolutional feature encoding layer, a multi-stage information fusion layer and a feature decoding layer.

[0029] The convolutional feature encoding layer uses a large convolution kernel to extract image shallow features of the to-be-processed image, to obtain time phase 1 shallow features and time phase 2 shallow features, and to perform splicing on the time phase 1 shallow features and the time phase 2 shallow features to obtain early-stage fusion features, and the time phase 1 shallow features, the time phase 2 shallow features and the early-stage fusion features are respectively input into a twin network to obtain time phase 1 feature maps, time phase 2 feature maps and early-stage fusion feature maps of the to-be-processed image through multi-layer convolution;

[0030] The multi-stage information fusion layer performs difference, addition and splicing on the time phase 1 feature maps and the time phase 2 feature maps at the same resolution, and splices the early-stage fusion feature maps after splicing, and performs information screening on a channel dimension through channel attention after splicing is completed, to obtain a middle-stage fusion feature map.

[0031] The feature decoding layer decodes the phase 1 feature map, the phase 2 feature map and the interim fusion feature map, obtains a phase 1 decoded feature map, a phase 2 decoded feature map and an interim fusion decoded feature map, and takes the phase 1 decoded feature map and the phase 2 decoded feature map as supplementary information, fuses the supplementary information with the interim fusion decoded feature map, and obtains a feature decoding map of the image to be processed.

[0032] Optionally, the convolutional neural network is an AlexNet network, a GoogLeNet network, a VGG network or a ResNe network.

[0033] The system of the present application can effectively avoid the problem of poor fitting effect of the training model caused by too many negative samples through the processing of the data enhancement module on the image. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The figure is a structural diagram of the system of the present application;

[0035] Figure 2 The figure is a flowchart of the system of the present application for detection;

[0036] Figure 3 The figure is a result display diagram of the present application in the LEVIR-CD dataset;

[0037] Figure 4 The figure is a flowchart of the method of the present application. DETAILED DESCRIPTION

[0038] Reference will now be made to the drawings to describe the exemplary embodiments of the present application in detail. The present application can be implemented in various forms, and is not limited to the embodiments described herein, which are provided to fully and completely disclose the present application and to fully convey the scope of the present application to those skilled in the art. The terms used in the exemplary embodiments represented in the drawings are not limited to the present application. In the drawings, the same elements / elements are denoted by the same reference numerals.

[0039] Unless otherwise defined, the terms used herein (including technical terms) have meanings commonly understood by those skilled in the art. In addition, it is to be understood that the terms defined in commonly used dictionaries should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art, and should not be interpreted in an idealized or overly formal sense.

[0040] The present application proposes a system 100 suitable for detecting changes in buildings, as shown in Figure 1 The system 100 comprises:

[0041] The data acquisition module 101 is configured to acquire a high-resolution remote sensing image of a target building, and use the high-resolution remote sensing image as a training set of an input data enhancement module;

[0042] The data enhancement module 102 is configured to train the training set, acquire a to-be-processed image, input the to-be-processed image into a convolutional neural network, pre-process the to-be-processed image through the convolutional neural network, and acquire a feature decoding image of the to-be-processed image.

[0043] The detection module 103 is configured to determine a change of the building according to the feature decoding image.

[0044] The convolutional neural network comprises a convolutional feature encoding layer, a multi-stage information fusion layer and a feature decoding layer.

[0045] The data enhancement module is configured to divide a large-size image into 2x2 or 3x3 sub-regions, randomly select a point in each divided region, and crop a candidate input image as a center of the selected point, and assign different candidate input images with probability weights according to the number of positive samples in the candidate image, so that the candidate input with a larger number of positive samples has a higher probability of being selected as an input.

[0046] The convolutional feature encoding layer is configured to use a large convolution kernel to extract shallow features to obtain a time phase 1 shallow feature and a time phase 2 shallow feature, and to splice the time phase 1 shallow feature and the time phase 2 shallow feature through a preliminary fusion layer to obtain a preliminary fusion feature. The time phase 1 shallow feature, the time phase 2 shallow feature and the preliminary fusion feature are input into a twin network, and a corresponding convolution feature map is obtained through multi-layer convolution: a time phase 1 feature map, a time phase 2 feature map and a preliminary fusion feature map.

[0047] The information fusion layer is configured to fuse the time phase 1 and the time phase 2 with the fusion feature, splice the time phase 1 and the time phase 2 features at the same resolution after difference, addition and splicing, and then splice the fusion feature, and then perform channel attention to filter information in the channel dimension to obtain a fused feature map.

[0048] The feature decoding layer is configured to input the time phase 1 feature map, the time phase 2 feature map and the intermediate fusion feature map into a convolutional feature decoding layer, and obtain corresponding decoding feature maps through multiple up-sampling convolutions: a time phase 1 decoding feature map, a time phase 2 decoding feature map and an intermediate fusion decoding feature map.

[0049] The convolutional neural network is an AlexNet network, a GoogLeNet network, a VGG network or a ResNe network.

[0050] As shown in Figure 2 The steps of the system 100 of the present application are as follows:

[0051] Step 1, data enhancement;

[0052] Firstly, the large size image is divided into 2x2 or 3x3 sub-regions, then a point is randomly selected in each segmented region, and the selected point is used as the center for cutting as a candidate input image, and different candidate input images are given probability weights according to the number of positive samples in the candidate image, so that the candidate input with more positive samples has a higher probability of being selected as the input.

[0053] Step 2, convolution feature encoding;

[0054] A large convolution kernel is used to extract shallow features to obtain phase 1 shallow features and phase 2 shallow features, and the phase 1 shallow features and the phase 2 shallow features are spliced through the early fusion layer to obtain the early fusion features. The phase 1 shallow features, the phase 2 shallow features and the early fusion features are respectively input into the twin network, and the corresponding convolution feature maps are obtained through multi-layer convolution: phase 1 feature map, phase 2 feature map and early fusion feature map.

[0055] Step 3, information fusion;

[0056] The phase 1 and the phase 2 are fused with the fusion features, the phase 1 and the phase 2 features at the same resolution are differentiated, added, spliced and then spliced with the fusion features, and then the channel attention is used for information screening in the channel dimension to obtain the fused feature map.

[0057] Step 4, feature decoding;

[0058] The phase 1 feature map, the phase 2 feature map and the intermediate fusion feature map are respectively input into the convolution feature decoding layer, and the corresponding decoding feature maps are obtained through multiple up-sampling convolutions: phase 1 decoding feature map, phase 2 decoding feature map and intermediate fusion decoding feature map. In the up-sampling process, the skip connection of the UNet network is used to transmit the information extracted by the feature extraction layer to the corresponding size feature map in the decoder to enhance the spatial details.

[0059] The effect of the present application can be further illustrated by the following simulation experiment.

[0060] 1. Simulation data set

[0061] To verify the effectiveness of the present application, the present application is used for experiment on the open source dataset LEVIR-CD proposed by Chen et al. in the literature "H. Chen and Z. Shi, A Spatial-Temporal Attention-Based Method and a New Dataset for Remote Sensing Image Change Detection, Remote Sensing, vol. 12, no. 10, pp. 1662-1684, 2020.", which contains 637 pairs of dual-time remote sensing images, of which the training set is 445 pairs, the verification set is 64 pairs, the test set is 128 pairs, the image resolution is 1024*1024 pixels, and the number of changed buildings is 31333.

[0062] 2. Simulation content

[0063] The present application achieves the best accuracy on the database, as shown in Table 1.

[0064] Table 1

[0065]

[0066] The present application also proposes a method suitable for detecting building changes, as shown in Figure 4 , comprising:

[0067] Step 1, collecting high-resolution remote sensing images of target buildings, and taking the high-resolution remote sensing images as a training set;

[0068] Step 2, training the training set, obtaining a to-be-processed image, and inputting the to-be-processed image into a convolutional neural network to pre-process the to-be-processed image through the convolutional neural network, and obtaining a feature decoding map of the to-be-processed image;

[0069] Step 3, determining the change of the building according to the feature decoding map.

[0070] Wherein, the training set is trained to obtain a to-be-processed image, comprising:

[0071] partitioning the high-resolution remote sensing images of the target buildings according to a preset size to obtain a plurality of sub-regions;

[0072] Randomly select an arbitrary point in the sub-region, and cut the sub-region with the arbitrary point as the center, and take the cut image as a candidate image;

[0073] Different probability weights are given to the candidate images according to the number of positive samples of the candidate images, so as to determine the probability of being selected, and the candidate image with a large probability of being selected is taken as a to-be-processed image.

[0074] The convolutional neural network comprises a convolutional feature encoding layer, a multi-stage information fusion layer and a feature decoding layer.

[0075] The convolutional feature encoding layer uses a large convolution kernel to extract image shallow features of the to-be-processed image, obtains a time phase 1 shallow feature and a time phase 2 shallow feature, and splices the time phase 1 shallow feature and the time phase 2 shallow feature to obtain a preliminary fusion feature.

[0076] The multi-stage information fusion layer splices the time phase 1 feature map and the time phase 2 feature map at the same resolution after difference, addition and splicing, and splices the preliminary fusion feature map, and after splicing is completed, information filtering is performed on the channel dimension through channel attention to obtain a medium-term fusion feature map.

[0077] The feature decoding layer decodes the time phase 1 feature map, the time phase 2 feature map and the medium-term fusion feature map to obtain a time phase 1 decoding feature map, a time phase 2 decoding feature map and a medium-term fusion decoding feature map, and takes the time phase 1 decoding feature map and the time phase 2 decoding feature map as supplementary information, and fuses the supplementary information with the medium-term fusion decoding feature map to obtain a feature decoding map of the to-be-processed image.

[0078] The convolutional neural network is an AlexNet network, a GoogLeNet network, a VGG network or a ResNe network.

[0079] The system has the characteristics of the multi-level information fusion layer, and can obtain experimental results by merging double-time phase image features of different levels.

[0080] 2. The present application proposes a positive and negative sample balancing algorithm. Since the remote sensing image widely exists the problem of sample imbalance, the research on the positive and negative sample balancing can effectively avoid that too many negative samples lead to poor fitting effect of the training model. Specifically:

[0081] The present application aims to balance the number of positive and negative samples when randomly clipping, for larger remote sensing images, first select multiple points distributed in different regions as candidate sample center points, calculate the number of positive samples in the candidate samples as a weight, and select the candidate samples according to the calculated weight to give different probabilities, so that the number of positive and negative samples is balanced under certain randomness, and the network can learn the change information more effectively.

[0082] The network framework of the present application aggregates information between different remote sensing images at different stages, the fusion module can fully learn the relationship between the double-time phase features, and the separate branch can retain the features of each time phase, and the fusion is carried out at different stages, so that the network can better learn the information of the double-time phase images.

[0083] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming languages Java and interpreted scripting language JavaScript.

[0084] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks

[0085] These computer program instructions can also be stored in a computer-readable storage medium that can guide the computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable storage medium produce a product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks

[0086] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart blocks Figure 1 The flowchart blocks

[0087] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to encompass within their scope all such variations and modifications as are included within the scope of the invention.

[0088] It is apparent that those skilled in the art can make modifications and variations to the application without departing from the spirit and scope of the application. Accordingly, it is intended to include all such modifications and variations in the scope of the claims.

Claims

1. A system suitable for detecting changes in a building, characterized in that, The system comprises: a data acquisition module for acquiring high-resolution remote sensing images of a target building and taking the high-resolution remote sensing images as a training set; a data enhancement module for expanding the training set, obtaining a to-be-processed image, inputting the to-be-processed image into a convolutional neural network, preprocessing the to-be-processed image through the convolutional neural network, and obtaining a feature decoding graph of the to-be-processed image; the data enhancement module expands the training set to obtain the to-be-processed image, including: partitioning the high-resolution remote sensing images of the target building according to a preset size to obtain a plurality of sub-regions; randomly selecting an arbitrary point in the sub-region and cropping the sub-region with the arbitrary point as the center to take the cropped image as a candidate image; assigning different probability weights to the candidate images according to the number of positive samples of the candidate images to determine the probability of the candidate images being selected, and taking the candidate images with a larger probability of being selected as the to-be-processed image; the convolutional neural network comprises a convolutional feature encoding layer, a multi-stage information fusion layer, and a feature decoding layer; the convolutional feature encoding layer uses a large convolution kernel to extract image shallow features of the to-be-processed image to obtain time phase 1 shallow features and time phase 2 shallow features, splices the time phase 1 shallow features and the time phase 2 shallow features to obtain early-stage fusion features, and inputs the time phase 1 shallow features, the time phase 2 shallow features, and the early-stage fusion features into a twin network respectively to obtain a time phase 1 feature map, a time phase 2 feature map, and an early-stage fusion feature map of the to-be-processed image through multiple layers of convolution; the multi-stage information fusion layer splices the time phase 1 feature map and the time phase 2 feature map at the same resolution after difference, addition, and splicing with the early-stage fusion feature map, performs information screening on the channel dimension through channel attention after splicing is completed, and obtains a middle-stage fusion feature map; the feature decoding layer decodes the time phase 1 feature map, the time phase 2 feature map, and the middle-stage fusion feature map to obtain a time phase 1 decoding feature map, a time phase 2 decoding feature map, and a middle-stage fusion decoding feature map, takes the time phase 1 decoding feature map and the time phase 2 decoding feature map as supplementary information, fuses the supplementary information with the middle-stage fusion decoding feature map, and obtains a feature decoding graph of the to-be-processed image; a detection module for determining a change of the building according to the feature decoding graph.

2. The system of claim 1, wherein, The convolutional neural network is an AlexNet network, a GoogLeNet network, a VGG network, or a ResNe network.

3. A method suitable for detecting changes in a building, characterized in that, The method comprises: acquiring high-resolution remote sensing images of a target building and taking the high-resolution remote sensing images as a training set; expanding the training set, obtaining a to-be-processed image, inputting the to-be-processed image into a convolutional neural network, preprocessing the to-be-processed image through the convolutional neural network, and obtaining a feature decoding graph of the to-be-processed image; the data enhancement module expands the training set to obtain the to-be-processed image, including: partitioning the high-resolution remote sensing images of the target building according to a preset size to obtain a plurality of sub-regions; Randomly selecting an arbitrary point in the sub-region, and cutting the sub-region with the arbitrary point as the center, taking the cut image as a candidate image; According to the number of positive samples of the candidate image, different probability weights are given to the candidate image to determine the probability of the candidate image being selected, and the candidate image with a larger selected probability is taken as a to-be-processed image; The convolutional neural network comprises a convolutional feature encoding layer, a multi-stage information fusion layer and a feature decoding layer. The convolutional feature encoding layer uses a large convolution kernel to extract image shallow features of the to-be-processed image, obtains a time phase 1 shallow feature and a time phase 2 shallow feature, splices the time phase 1 shallow feature and the time phase 2 shallow feature to obtain a preliminary fusion feature, inputs the time phase 1 shallow feature, the time phase 2 shallow feature and the preliminary fusion feature into a twin network respectively, and obtains a time phase 1 feature map, a time phase 2 feature map and a preliminary fusion feature map of the to-be-processed image through multi-layer convolution. The multi-stage information fusion layer splices the time phase 1 feature map and the time phase 2 feature map at the same resolution after difference, addition and splicing, and splices the preliminary fusion feature map, after the splicing is completed, information screening in the channel dimension is performed through channel attention to obtain a medium-term fusion feature map. The feature decoding layer decodes the time phase 1 feature map, the time phase 2 feature map and the medium-term fusion feature map to obtain a time phase 1 decoding feature map, a time phase 2 decoding feature map and a medium-term fusion decoding feature map, takes the time phase 1 decoding feature map and the time phase 2 decoding feature map as supplementary information, fuses the supplementary information with the medium-term fusion decoding feature map, and obtains a feature decoding map of the to-be-processed image. According to the feature decoding map, a change of the building is determined.

4. The method of claim 3, wherein, The convolutional neural network is an AlexNet network, a GoogLeNet network, a VGG network or a ResNe network.

Images