Image recognition method and system based on combination of segmentation and color

Abstract

The present application provides an image recognition method and system based on the combination of segmentation and color, the method includes obtaining an image to be processed, then preprocessing the image to be processed to obtain a preprocessed image, extracting features from the preprocessed image, and performing normalization to obtain a feature-extracted image, further, processing the acquired feature-extracted image by establishing a principal component evaluation model to generate a feature-processed image, the principal component evaluation model is established based on the HSV color model; finally, the feature-processed image is input into a preset residual network model based on a multi-scale segmentation algorithm to generate a recognition result. The present application adopts an image recognition technology that combines a multi-scale segmentation algorithm with an HSV color model, which can effectively monitor and identify common intrusive foreign objects such as drones and kites in special places such as substations and transmission lines, solving the problem of relying on artificially designed features in traditional image processing and difficulty in extracting deep information.

Description

Technical Field

[0001] The present application relates to the technical field of image recognition, and in particular to an image recognition method and system based on the combination of segmentation and color. Background Art

[0002] With economic growth and an increase in urban population, power companies must ensure the normal use of electricity in people's lives and work, and be responsible for the quality and safety of electricity use. The safe operation of the power grid is the basic guarantee for the power supply of the power system. Safe production and transmission are the core requirements of power safety. At the power production and transmission site, obstacles that affect safety appear, especially drones and kites. In an environment where equipment is running at high speed, even small foreign objects can cause very serious accidents. For example, these intrusions may cause line failures, and even cause local or even regional power outages.

[0003] In order to ensure the safety of power production and transmission sites and enable the power environment to operate safely, reliably, efficiently and stably, it is necessary to effectively monitor the intrusion of foreign objects, detect intrusive foreign objects in a timely manner, accurately understand the status of substations, and discover dangerous situations as early as possible.

[0004] However, traditional image processing has always relied on artificially designed features, which only describe and characterize low-level edge information in the image and are difficult to extract deep information. The current foreign body detection method can solve the problem of intrusion foreign body detection to a certain extent, but the detection speed and accuracy are not high, and it cannot effectively detect intrusion foreign bodies in real time. Summary of the invention

[0005] The present application provides an image recognition method and system based on the combination of segmentation and color to solve the problem that traditional image processing relies on manually designed features, which can only describe and characterize low-level edge information in the image and are difficult to extract deep-level information.

[0006] The technical solutions adopted by this application to solve the above technical problems are as follows:

[0007] An image recognition method based on combining segmentation and color, comprising the following steps:

[0008] Get the image to be processed;

[0009] Acquire a preprocessed image, wherein the preprocessed image is obtained by preprocessing the image to be processed;

[0010] Acquire a feature extraction image, wherein the feature extraction image is obtained by performing feature extraction on the preprocessed image and performing normalization processing;

[0011] Generate a feature-processed image, wherein the feature-processed image is obtained by processing the acquired feature-extracted image by establishing a principal component evaluation model, wherein the principal component evaluation model is established based on the HSV color model;

[0012] Generate a recognition result, wherein the recognition result is obtained by inputting the feature-processed image into a preset residual network model based on a multi-scale segmentation algorithm.

[0013] Optionally, obtaining the image to be processed includes:

[0014] By dynamically configuring the RF parameters in real time, different types of images to be processed can be obtained at different times.

[0015] Optionally, the obtaining of a preprocessed image includes:

[0016] Perform image denoising and image enhancement processing on the image to be processed.

[0017] Optionally, acquiring the feature extraction image includes:

[0018] The features of the image similarity and gradient, color moment, peak signal-to-noise ratio, mean variance and number of pixels in the obtained preprocessed image are extracted and normalized to obtain a feature-extracted image.

[0019] Optionally, acquiring the feature extraction image includes:

[0020] Feature extraction is performed on 10 indicators including value, variance, signal-to-noise ratio, similarity, horizontal gradient, vertical gradient, number of pixels, first-order moment, second-order moment and third-order moment of the preprocessed image;

[0021] Normalize the extracted features.

[0022] Optionally, generating a feature-processed image includes:

[0023] Get the RGB format image of the image to be detected;

[0024] Convert the RGB format image into an HSV image;

[0025] Based on the HSV color model, a principal component evaluation model is established;

[0026] The extracted features are processed by establishing a principal component evaluation model to generate a feature-processed image.

[0027] Optionally, before generating the recognition result, the method further includes:

[0028] The multi-scale segmentation method is improved by reducing the number of deep convolutional network layers and increasing the network width, and an image recognition residual network based on multi-scale segmentation is obtained.

[0029] Optionally, the method of improving the multi-scale segmentation method by reducing the number of deep convolutional network layers and increasing the network width includes:

[0030] ResNet is improved by combining the residual structure of ResNet and the method of increasing network width by WRN;

[0031] By reducing the number of network layers, a network model with only 7 layers was obtained on the cifar dataset;

[0032] Compare and calculate different improved modules, and replace the 3×3 convolution in ResNet with the multi-scale segmentation module SS-Block;

[0033] The multi-scale segmentation network model with two 1×1 dimensionality reduction and dimensionality increase convolutions in the ResNet bottleneck module is discarded to obtain an image recognition residual network based on multi-scale segmentation.

[0034] An image recognition system based on the combination of segmentation and color, comprising:

[0035] The image to be processed acquiring unit is used to acquire the image to be processed;

[0036] A preprocessed image acquisition unit, used for acquiring a preprocessed image, wherein the preprocessed image is obtained by preprocessing the image to be processed;

[0037] A feature extraction image acquisition unit, used to acquire a feature extraction image, wherein the feature extraction image is obtained by performing feature extraction on the preprocessed image and performing normalization processing;

[0038] A feature processing image generating unit, used for generating a feature processing image, wherein the feature processing image is obtained by processing the acquired feature extraction image by establishing a principal component evaluation model, wherein the principal component evaluation model is established based on the HSV color model;

[0039] The recognition result generating unit is used to generate a recognition result, wherein the recognition result is obtained by inputting the feature-processed image into a preset residual network model based on a multi-scale segmentation algorithm.

[0040] The technical solution provided by this application includes the following beneficial technical effects:

[0041] The present application provides an image recognition method and system based on the combination of segmentation and color, the method comprising obtaining an image to be processed, then preprocessing the image to be processed to obtain a preprocessed image, extracting features from the preprocessed image, and performing normalization to obtain a feature-extracted image, further, processing the acquired feature-extracted image by establishing a principal component evaluation model to generate a feature-processed image, the principal component evaluation model is established based on the HSV color model; finally, the feature-processed image is input into a preset residual network model based on a multi-scale segmentation algorithm to generate a recognition result. The present application adopts an image recognition technology that combines a multi-scale segmentation algorithm with an HSV color model, which can effectively monitor and identify common intrusive foreign objects such as drones and kites in special places such as substations and transmission lines, solving the problem of relying on artificially designed features in traditional image processing, and the feature can only describe and characterize low-level edge information in the image, and it is difficult to extract deep-level information. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0043] Figure 1 A flow chart of an image recognition method based on the combination of segmentation and color provided in an embodiment of the present application;

[0044] Figure 2 A schematic diagram of the convolution process of the SS-Block provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the application embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of this application; it is obvious that the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0046] See also Figure 1 The image recognition method based on the combination of segmentation and color provided in the embodiment of the present application includes the following steps:

[0047] S1. Obtain an image to be processed.

[0048] The image to be processed is a target image, that is, an image obtained by automatically monitoring the intrusion of foreign objects. Such images need to be processed by foreign object detection because of the intrusion of foreign objects.

[0049] In some embodiments, the image to be processed can be obtained by dynamically configuring the radio frequency parameters in real time to obtain different types of images to be processed at different times. While obtaining the detection image, the radio frequency parameters are continuously dynamically configured to adapt to different types of drone detection and identification technologies, and to specifically judge the intrusion of foreign objects such as drones that have signal transmission and reception types, and then transmit the image at the receiving signal to the image recognition system.

[0050] S2. Acquire a preprocessed image, where the preprocessed image is obtained by preprocessing the image to be processed.

[0051] First, it is necessary to collect the foreign body intrusion limit images obtained by monitoring. After the collection, the image to be processed is determined, and then the image to be processed is preprocessed. The preprocessing method includes image denoising and image enhancement processing on the image to be processed.

[0052] Due to the influence of different environmental factors such as weather and temperature, the images obtained by the surveillance camera have completely different imaging characteristics, such as low resolution, low saturation, reduced contrast, and different degrees of noise. These characteristics affect each process of foreign body intrusion detection. Before inputting these images under different scene conditions into the intrusion foreign body detection system, these images must be preprocessed as necessary. The main methods used for image preprocessing in the embodiments of this application are image denoising and image enhancement.

[0053] Image denoising refers to the process of reducing noise in digital images. First, the speckle noise and salt and pepper noise in the image to be processed are processed by median filtering. Median filtering is the result of taking an odd number of data from a window in the image and arranging them. The median refers to the number located in the center of the 3*3 window after arranging them in descending order; median filtering is to perform median processing on all pixel values ​​in the window.

[0054] R=medium(Z i )i=1~9

[0055] Among them, R is the pixel value replaced after the median filter of the coordinate point to be processed, and Z is the pixel value of each pixel in the window. Median filtering is particularly effective in eliminating pulse interference and isolated noise and processing the blur of image details; image enhancement uses histogram equalization to achieve image enhancement. For some images with grayscale concentrated in a narrow interval and lacking detailed information, histogram correction can be used to widen the grayscale spacing of the image or make the grayscale distribution more uniform, thereby improving the contrast of the image and achieving the purpose of image enhancement.

[0056] S3. Acquire a feature extraction image, wherein the feature extraction image is obtained by performing feature extraction on the preprocessed image and performing normalization processing.

[0057] By extracting features from the preprocessed image, a principal component evaluation model is established to obtain indicators. Then, feature extraction is performed based on multiple features of the image, such as color moment, peak signal-to-noise ratio, structural similarity index, etc. After normalization, a principal component analysis model is established to process five major indicators (similarity and gradient, color moment, peak signal-to-noise ratio, mean variance and number of pixels) to prevent image feature loss.

[0058] As an implementation manner, the acquiring of the feature extraction image includes:

[0059] Feature extraction is performed on 10 indicators including value, variance, signal-to-noise ratio, similarity, horizontal gradient, vertical gradient, number of pixels, first-order moment, second-order moment and third-order moment of the preprocessed image;

[0060] Normalize the extracted features.

[0061] S4. Generate a feature-processed image, wherein the feature-processed image is obtained by processing the acquired feature-extracted image through establishing a principal component evaluation model, and the principal component evaluation model is established based on the HSV color model.

[0062] Since important information is often lost during image preprocessing, the HSV color model can be used to retain the image's texture information and global color information to a great extent when extracting features, making it less likely to cause image feature loss. Especially for kite foreign bodies, due to their bright colors and regular shapes, the feature retention rate is much higher than that of traditional feature extraction methods.

[0063] When implementing the operation, it is as follows:

[0064] The preprocessed image is input into a trained image recognition model based on the combination of segmentation and color, and the detection result is output; wherein, the color model is an HSV color model, and the HSV color model uses the hue, saturation, and brightness of the HSV color space to be relatively consistent with the subjective perception of color by the human eye, especially the kite, which is usually a foreign body with a more distinct color feature. Compared with other color spaces, the HSV space can better reflect human perception of color, so the HSV color model is adopted; wherein, the segmentation method is a multi-scale segmentation method. For the residual network, one of the deep convolutional networks, increasing the network depth and width can generate rich feature information, and the Hierarchical-SplitBlock (HS-Block) obtained by the multi-scale segmentation method is used to improve the performance of the existing neural network. HS-Block contains multiple connection and merging operations, which together constitute the multi-scale feature extraction of the Block; at the same time, HS-Block has better flexibility and efficiency, can make full use of feature information, and reduce redundant feature information, so multi-scale segmentation of the residual structure can effectively improve the recognition accuracy.

[0065] Since the color of a kite can be expressed in many ways, the HSV color model is a model suitable for naked eye discrimination. It represents color information as three attributes, namely hue h, saturation s and brightness v. This color model can be expressed in a three-dimensional space coordinate system, and the conversion formula between it and the RGB space is:

[0066]

[0067]

[0068]

[0069] For brightly colored foreign objects such as kites, with the sky as the pure background, the hue h and saturation s are relatively high; for intrusive foreign objects such as drones, their own light reflection is relatively large, which is well reflected in the brightness v.

[0070] First, convert the RGB format surveillance image to be detected into an HSV image; then define an object of Mat type, that is, the mask; then define the upper and lower limits of the color. The upper limit is a Scalar object containing three values: h min ,s min ,v min Indicates the minimum value of the three elements of HSV, and the lower limit is also a Scalar object containing three values: h max ,s max , v maxIndicates the maximum value of the three HSV elements; finally, use the inRange function to detect whether each pixel of the src image is in the lower (h min ,s min ,v min ) and upper(h max ,s max , v max ), if so, the pixel is set to 255 and saved in the mask image, otherwise it is 0. Finally, the region of interest image is saved as the input of the subsequent network.

[0071] S5. Generate a recognition result, wherein the recognition result is obtained by inputting the feature-processed image into a preset residual network model based on a multi-scale segmentation algorithm.

[0072] The detection image and the extracted features obtained above are sent to the image recognition residual network based on multi-scale segmentation. The recognition speed is greatly improved due to the shortened network length and the multi-scale segmentation method is also able to greatly improve the recognition accuracy. In particular, combined with the radio signal detection technology in step S1, the recognition accuracy of two common intrusion foreign objects, drones and kites, can reach a high level.

[0073] When implementing the operation, it is as follows:

[0074] In an embodiment of the present application, the multi-scale segmentation method proposed by HS-ResNet is improved, and a simple multi-scale segmentation method is proposed. By reducing the number of layers of the deep convolutional network and increasing the width, the recognition speed is made faster while ensuring that the recognition accuracy and parameter amount of the network are similar.

[0075] First, ResNet is improved by combining the residual structure of ResNet and the WRN method of increasing the network width. Then, by reducing the number of network layers, a network model with a network length of only 7 layers is obtained on the cifar dataset to ensure recognition accuracy. Finally, the multi-scale segmentation method of HS-ResNet is improved. The original method can only make feature information of different groups enjoy receptive fields of different scales, so that in the feature information connected in the front, the number of convolutions is small, the receptive field is small, and more attention is paid to detail information; in the feature information connected in the back, the number of convolutions is large, the receptive field is large, and more attention is paid to global information. Through receptive fields of different sizes, the richness of feature information is increased.

[0076] Secondly, replace the 3×3 convolution in ResNet with a new multi-scale segmentation module SS-Block. When the feature information is input, the first step is to divide the feature information input by the 1×1 convolution into two equal parts according to the number of channels (if the number of channels is an odd number, round it down); the second step, half of the feature information is directly sent to the end for merging, and the other half of the feature information is convolved; the third step is to repeat the above steps until the last feature information; the fourth step is to merge the last feature information with the first half of the feature information obtained previously, and output them together to the 1×1 convolution. The multi-scale segmentation module of SSRNet is similar to the bottleneck module structure of ResNet, except that the 3×3 convolution in the bottleneck module of ResNet is replaced by SS-Block. Among them, the convolution of SS-Block is expressed as 3×3 convolution + batch normalization + Relu activation function. Figure 2 shown.

[0077] Since the connection merging operation fuses different features and enhances the information interaction of different groups, it requires a certain amount of work. Therefore, on the basis of improving ResNet, the multiple connection merging operations in the multi-scale segmentation method of HS-ResNet are optimized, and only the last connection merging operation is retained, that is, only one superposition of feature information and one merging of channels are performed, which greatly improves the recognition speed, and finally obtains the image recognition residual network (SSRNet) based on multi-scale segmentation.

[0078] As the network length increases, the receptive field of the convolutional neural network increases. However, not all receptive fields contribute equally to the output features. The central area of ​​the receptive field has a greater impact on the output features. Through experiments, it was found that when the network length is set to 1.5 to 2 times the ratio of the receptive field size to the input image size, the network has a higher recognition accuracy and a faster recognition speed, achieving a better balance between accuracy and speed. The calculation formula for the k-th layer receptive field is as follows:

[0079]

[0080] In the formula, l k-1 is the receptive field size of the k-1th layer, f k is the convolution kernel size of the current layer, s i is the stride of the i-th layer.

[0081] The multi-scale segmentation network model SSRNet proposed in the embodiment of the present application replaces the 3×3 convolution in ResNet with a new multi-scale segmentation module SS-Block. According to the above formula, different improved modules are compared and calculated. Finally, a multi-scale segmentation network model is adopted that replaces the 3×3 convolution in the bottleneck module of ResNet and removes the two 1×1 dimensionality reduction and dimensionality increase convolutions in the bottleneck module of ResNet. The receptive field size of the final output of the model is exactly 1.5 to 2 times the size of the input image after 3 scale segmentations. Through training, the recognition speed of the model is acceptable, and the recognition error rate is 4.04%. The length of the network model can achieve a good balance between recognition accuracy and recognition speed.

[0082] Corresponding to the above-mentioned embodiment of the image recognition method based on the combination of segmentation and color, the present application also provides an embodiment of an image recognition system based on the combination of segmentation and color. The system comprises:

[0083] The image to be processed acquiring unit is used to acquire the image to be processed;

[0084] A preprocessed image acquisition unit, used for acquiring a preprocessed image, wherein the preprocessed image is obtained by preprocessing the image to be processed;

[0085] A feature extraction image acquisition unit, used to acquire a feature extraction image, wherein the feature extraction image is obtained by performing feature extraction on the preprocessed image and performing normalization processing;

[0086] A feature processing image generating unit, used for generating a feature processing image, wherein the feature processing image is obtained by processing the acquired feature extraction image by establishing a principal component evaluation model, wherein the principal component evaluation model is established based on the HSV color model;

[0087] The recognition result generating unit is used to generate a recognition result, wherein the recognition result is obtained by inputting the feature-processed image into a preset residual network model based on a multi-scale segmentation algorithm.

[0088] In summary, the image recognition method and system based on the combination of segmentation and color provided by the embodiment of the present application include obtaining an image to be processed, then preprocessing the image to be processed to obtain a preprocessed image, extracting features from the preprocessed image, and performing normalization processing to obtain a feature extraction image. Furthermore, a principal component evaluation model is established to process the acquired feature extraction image to generate a feature processing image, and the principal component evaluation model is established based on the HSV color model; finally, the feature processing image is input into a preset residual network model based on a multi-scale segmentation algorithm to generate a recognition result. The image recognition technology combining a multi-scale segmentation algorithm with an HSV color model is adopted in the present application, which can effectively monitor and identify common intrusive foreign objects such as drones and kites in special places such as substations and transmission lines, and solves the problem that traditional image processing relies on artificially designed features, and the feature can only describe and characterize low-level edge information in the image, and it is difficult to extract deep information.

[0089] It should be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0090] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features disclosed herein.

[0091] It should be understood that the present application is not limited to what has been described above and shown in the drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. An image recognition method based on the combination of segmentation and color, characterized in that: The following steps are involved: Get the image to be processed; Acquire a preprocessed image, wherein the preprocessed image is obtained by preprocessing the image to be processed; Acquire a feature extraction image, wherein the feature extraction image is obtained by performing feature extraction on the preprocessed image and performing normalization processing; The step of acquiring a feature extraction image comprises: Extract features of the image similarity and gradient, color moment, peak signal-to-noise ratio, mean variance and number of pixels from the acquired preprocessed image, and perform normalization to obtain a feature-extracted image; The acquiring of the feature extraction image further comprises: Feature extraction is performed on 10 indicators including value, variance, signal-to-noise ratio, similarity, horizontal gradient, vertical gradient, number of pixels, first-order moment, second-order moment and third-order moment of the preprocessed image; Normalize the extracted features; Generate a feature-processed image, wherein the feature-processed image is obtained by processing the acquired feature-extracted image by establishing a principal component evaluation model, wherein the principal component evaluation model is established based on the HSV color model; The generating of the feature-processed image comprises: Get the RGB format image of the image to be detected; Convert the RGB format image into an HSV image; Based on the HSV color model, a principal component evaluation model is established; The extracted features are processed by establishing a principal component evaluation model to generate a feature-processed image; Generate a recognition result, wherein the recognition result is obtained by inputting the feature-processed image into a preset residual network model based on a multi-scale segmentation algorithm, wherein the multi-scale segmentation algorithm is an improvement of the HS-ResNet multi-scale segmentation method.

2. The image recognition method based on the combination of segmentation and color according to claim 1, characterized in that: The step of obtaining the image to be processed comprises: By dynamically configuring the RF parameters in real time, different types of images to be processed can be obtained at different times.

3. The image recognition method based on the combination of segmentation and color according to claim 1, characterized in that: The obtaining of the preprocessed image comprises: Perform image denoising and image enhancement processing on the image to be processed.

4. The image recognition method based on the combination of segmentation and color according to claim 1, characterized in that: Before generating the recognition result, the method further includes: The multi-scale segmentation method is improved by reducing the number of deep convolutional network layers and increasing the network width, and an image recognition residual network based on multi-scale segmentation is obtained.

5. The image recognition method based on the combination of segmentation and color according to claim 4, characterized in that: The multi-scale segmentation method is improved by reducing the number of deep convolutional network layers and increasing the network width, including: ResNet is improved by combining the residual structure of ResNet and the method of increasing the network width by WRN; By reducing the number of network layers, a network model with only 7 layers was obtained on the cifar dataset; Compare and calculate different improved modules, and replace the 3×3 convolution in ResNet with the multi-scale segmentation module SS-Block; the convolution of SS-Block is represented by 3×3 convolution + batch normalization + Relu activation function; The multi-scale segmentation network model with two 1×1 dimensionality reduction and dimensionality increase convolutions in the ResNet bottleneck module is discarded to obtain an image recognition residual network based on multi-scale segmentation.

6. An image recognition system based on the combination of segmentation and color, characterized in that: include: The image to be processed acquiring unit is used to acquire the image to be processed; A preprocessed image acquisition unit, used for acquiring a preprocessed image, wherein the preprocessed image is obtained by preprocessing the image to be processed; A feature extraction image acquisition unit, used to acquire a feature extraction image, wherein the feature extraction image is obtained by performing feature extraction on the preprocessed image and performing normalization processing; The step of acquiring a feature extraction image comprises: Extract features of the image similarity and gradient, color moment, peak signal-to-noise ratio, mean variance and number of pixels from the acquired preprocessed image, and perform normalization to obtain a feature-extracted image; The acquiring of the feature extraction image further comprises: Feature extraction is performed on 10 indicators including value, variance, signal-to-noise ratio, similarity, horizontal gradient, vertical gradient, number of pixels, first-order moment, second-order moment and third-order moment of the preprocessed image; Normalize the extracted features; A feature processing image generating unit, used for generating a feature processing image, wherein the feature processing image is obtained by processing the acquired feature extraction image by establishing a principal component evaluation model, wherein the principal component evaluation model is established based on the HSV color model; The generating of the feature-processed image comprises: Get the RGB format image of the image to be detected; Convert the RGB format image into an HSV image; Based on the HSV color model, a principal component evaluation model is established; The extracted features are processed by establishing a principal component evaluation model to generate a feature-processed image; The recognition result generating unit is used to generate a recognition result, wherein the recognition result is obtained by inputting the feature-processed image into a preset residual network model based on a multi-scale segmentation algorithm, wherein the multi-scale segmentation algorithm is an improvement of the HS-ResNet multi-scale segmentation method.

Images