WO2017132933A1

WO2017132933A1 - Image processing method and related apparatus

Info

Publication number: WO2017132933A1
Application number: PCT/CN2016/073473
Authority: WO
Inventors: 汪涛; 姚骏; 柴振华
Original assignee: 华为技术有限公司
Priority date: 2016-02-04
Filing date: 2016-02-04
Publication date: 2017-08-10
Also published as: CN107735800B; CN107735800A

Abstract

An image processing method, an image processing apparatus and a computing device. The method comprises: acquiring a target image (201), determining, by means of a classifier, from N partition models, a target partition model that the target image corresponds to (202), and then partitioning the target image using the target partition model, so as to obtain the partition result of the target image (203). The image processing apparatus uses the classifier to select the appropriate partition model in N partition models instead of using a single partition model, and this ensures that the image processing device is able to select, for each type of image, the partition model which is appropriate in terms of both accuracy and speed so as to partition the target image, being able to meet the actual needs.

Description

Image processing method and related device

Technical field

The present invention relates to the field of image processing, and in particular, to an image processing method and related apparatus.

Background technique

In the field of image processing, the image processing apparatus generally realizes the detection and recognition of the target in the image through two steps of partitioning and classification, as shown in FIG. 1: the image partition model accepts the input image, and divides the input image into regions of different sizes. The image classification model uses a convolutional neural network or other classification algorithm to continuously extract the features of each region of the image through a hierarchical structure, and finally identify the target object.

In practical applications, the size, composition, complexity, etc. of images are ever-changing, and different images are suitable for partitioning with different partitioning algorithms. Therefore, the current technology has developed a variety of image partition models for different kinds of images. Among them, different image partitioning models adopt different image partitioning algorithms, so when partitioning the same sub-image, the accuracy of the obtained partitioning result and the duration of the partitioning operation are different. However, in general, the same image partition model cannot take both precision and speed into consideration. The image partition model with higher precision tends to have a slower partitioning speed and a longer partitioning operation. The image partitioning model with faster partitioning speed tends to be less accurate. Therefore, the image partitioning model with higher precision is suitable for processing more complex images to increase the credibility of image partitioning; the image partitioning model with lower precision is suitable for processing relatively simple images to speed up image partitioning.

The image partitioning models that are suitable for different images are different. Therefore, the image processing apparatus cannot guarantee that all images are suitable regardless of whether the image partitioning model with high precision or low precision is used. This results in the current image processing device not being able to balance the accuracy and speed of image processing, and cannot meet the actual needs.

Summary of the invention

The invention provides an image processing method and device for partitioning a target image by using a suitable image partitioning model.

A first aspect of the present invention provides an image processing method suitable for use in an image processing apparatus. The image processing device acquires a target image to be processed, and then uses a classifier to determine a target partition model corresponding to the target image from the N partition models, where N is an integer greater than 1. Among them, the classifier can It is a classification function or classification model, and can also be a hardware unit with processing functions for mapping a given image onto one of the N partition models. The image processing device inputs the target image into the classifier, and can determine the target partition model based on the output of the classifier. After the target partition model is determined, the image processing apparatus partitions the target image using the target partition model to obtain a partition result of the target image. In the image processing method provided by the present invention, the image processing apparatus does not adopt a single partition model, but uses a classifier to select an appropriate partition model in the N partition models, thereby ensuring that the image processing apparatus can perform all types of images. A partition model with appropriate accuracy and speed is selected to partition the target image to meet actual needs.

Optionally, the image processing apparatus may also use the first partition model to partition the target image in parallel while determining the target partition model using the classifier. Wherein, the time for determining the target partition model according to the classifier is shorter than the time for partitioning the target image by using the first partition model, so that if the image processing apparatus determines that the target partition model is the first partition model, the first partition model is continuously used. Partitioning the target image is equivalent to the image processing apparatus directly zoning the target image using the first partition model after acquiring the target image, which can save time for using the classifier to determine the target partition model. If the image processing apparatus determines that the target partition model is not the first partition model, the image processing apparatus stops the operation of partitioning the target image using the first partition model, and re-partitions the target image using the target partition model to obtain a partition of the target image. result. The first partition model may be the highest-precision partition model of the N partition models, or the partition model with the longest time for performing image partitioning operations, or a partition model determined by other standards.

Optionally, the classifier may be stored locally by the image processing device and read locally from the use; or may be obtained from other places by the image processing device; or may be imaged by the image processing device through the neural network. The operation is obtained.

Optionally, the image processing apparatus may train the classifier by: acquiring P training images in advance; using the N partition models for each training image, and determining each of the partition models for each training image. The ratio of the accuracy of the partition to the time is performed; then the partition model with the largest ratio of the precision of the partitioning of the i-th training image to the time is determined as the partition model corresponding to the i-th image, 1≤i≤P. After determining the partition model corresponding to each training image, the error back propagation (abbreviation: BP) algorithm or other algorithms are used to obtain the classifier. As long as the P training images cover various types of images, and the P value is large enough, the classification can be guaranteed. The device has a classification effect that meets the actual needs. The classifier obtained by such a method can determine, as the target partition model of the target image, the partition model in which the ratio of the accuracy of the partitioning of the target image and the time is the largest among the N partition models.

Optionally, after obtaining the partition result of the target image, the graphics processing apparatus may further classify the partition result of the target image. The image processing apparatus may determine the target size of the convolution kernel of the corresponding classification model according to the accuracy of the target partition model, and then use the convolution kernel of the target size to classify the partition result of the target image to obtain the target image. Classification results. Such a method can classify more complex images by a larger convolution kernel, and a simpler image is classified by a smaller convolution kernel, taking into account the accuracy and speed of the classification calculation.

Optionally, the image processing apparatus may further determine a calculation channel of the corresponding classification model according to the accuracy of the target partition model. Specifically, if the target partition model is the most accurate partition model among the N partition models, the image processing apparatus can use the floating point channel and the fixed point channel to classify the partitioning result of the target image in parallel to speed up the classification operation. Improve the efficiency of image classification. If the target partition model is not the most accurate model among the N partition models, the image processing apparatus may classify the partition result of the target image using only the floating point channel to obtain the classification result of the target image.

A second aspect of the present invention provides an image processing apparatus including an image acquisition module, a model determination module, and an image recognition module. The image obtaining module is configured to acquire a target image to be processed; the model determining module is configured to determine, in the N partition models, a target partition model corresponding to the target image by using a classifier, where N is an integer greater than 1; and the image recognition module uses The image is identified, and the target image is partitioned by using the target partition model to obtain a partition result of the target image.

Optionally, the image recognition module is specifically configured to: when the target partition model is determined by using the classifier, the first partition model may be used in parallel to partition the target image. The first partition model may be the highest-precision partition model of the N partition models, or the partition model with the longest time for performing image partitioning operations, or a partition model determined by other standards. If the image processing apparatus determines that the target partition model is the first partition model, then continuing to partition the target image by using the first partition model, which is equivalent to the image processing apparatus directly partitioning the target image by using the first partition model after acquiring the target image. , you can save time using the classifier to determine the target partition model. If the image processing apparatus determines that the target partition model is not the first partition model, the image processing apparatus stops the operation of partitioning the target image using the first partition model, and re-partitions the target image using the target partition model. Get the partition result of the target image. The first partition model is the partition model with the highest precision among the N partition models or the partition model with the longest time for image partitioning operation.

Optionally, the model determining module is further configured to: obtain a classifier by using an image training operation of the neural network.

Optionally, the model determining module may train the classifier by: acquiring P training images in advance; using the N partition models for each training image, and determining each of the partition models for each training image. The ratio of the accuracy of the partition to the time is performed; then the partition model with the largest ratio of the precision of the partitioning of the i-th training image to the time is determined as the partition model corresponding to the i-th image, 1≤i≤P. After determining the partition model corresponding to each training image, the BP algorithm or other algorithms are used to obtain the classifier.

Optionally, the image recognition module is further configured to: determine a target size of the convolution kernel of the corresponding classification model according to the precision of the target partition model, and then use the convolution kernel of the target size to perform the partitioning result of the target image. Classification, and the classification result of the target image is obtained.

Optionally, the image recognition module is further configured to: determine a calculation channel of the corresponding classification model according to the accuracy of the target partition model. If the target partition model is the most accurate partition model among the N partition models, the partitioning results of the target image are classified in parallel using the floating point channel and the fixed point channel. If the target partition model is not the most accurate model among the N partition models, the partition result of the target image is classified using only the floating point channel, and the classification result of the target image is obtained.

A third aspect of the invention provides a computer apparatus comprising a processor, a memory and a communication interface for performing the image processing method provided by the first aspect of the invention by invoking instructions in the memory.

In the image processing method provided by the present invention, the image processing apparatus acquires a target image, determines a target partition model corresponding to the target image in the N partition models through a classifier, and then partitions the target image using the target partition model to obtain a partition of the target image. result. The image processing apparatus of the present invention does not adopt a single partition model, but uses a classifier to select an appropriate partition model in the N partition models, thereby ensuring that the image processing apparatus can select accuracy and speed for all types of images. A more suitable partition model to partition the target image can meet the actual needs.

DRAWINGS

1 is a schematic diagram showing the principle of detecting and recognizing an image by an image processing apparatus;

2 is a flowchart of an embodiment of an image processing method according to an embodiment of the present invention;

FIG. 3 is a structural diagram of an embodiment of an image processing apparatus according to an embodiment of the present invention; FIG.

FIG. 4 is a structural diagram of an embodiment of a computing device according to an embodiment of the present invention.

detailed description

The present invention provides an image processing method and apparatus, which will be separately described below.

In practical applications, the size, composition, complexity, etc. of images are ever-changing, and different images are suitable for partitioning with different partitioning algorithms. So at this stage the technology has developed multiple partition models for different kinds of images. Among them, different partition models use different image partitioning algorithms, so when partitioning the same sub-image, the accuracy of the obtained partitioning result and the duration of the partitioning operation are different. However, in general, the same partition model cannot take into account both accuracy and speed. The partition model with higher precision tends to have slower partition speed and longer partition operation. The partition model with fast partition speed tends to be less accurate. Therefore, the higher precision partition model is suitable for processing more complex images to increase the credibility of image partitioning; the lower precision partition model is suitable for processing simpler images to speed up image partitioning.

In the current technology, image processing devices often use a single fixed partition model, and a single partition model cannot meet the accuracy and speed requirements of all images. To this end, the present invention provides an image processing method capable of adopting different partition models for different images, and the basic flow thereof is shown in FIG. 2, including

201. Acquire a target image to be processed.

202. Determine, by using a classifier, a target partition model corresponding to the target image.

In the present invention, the image processing apparatus may determine a target partition model corresponding to the target image from the N partition models. The N partition models may be stored locally by the image processing device and may be locally read during use; or may be obtained from other places by the image processing device, which is not limited herein. Where N is an integer greater than one.

The determined target partition model should be the partition model of the N partition models that is most suitable for processing the target image. There are many methods for measuring which partition model of the N partition models is most suitable for processing the target image. For example, a partition model with the largest ratio of precision to time for partitioning the target image can be considered as the most suitable for processing the target image. Partition model.

Among them, there are many methods for determining the target partition model. In the present invention, a classifier (English: classifier) is used to determine the target partition model. The essence of the classifier can be a classification function or a classification model, which can classify a given image sample into N classes, each class corresponding to a partition model, thereby mapping the image into the N partition models. The functionality of a partition model. The image classification device inputs the target image to the classifier, and can determine the target partition model based on the output of the classifier. Wherein, the classifier can be stored locally by the image processing device and read locally from use; it can also be obtained from other places by the image processing device; or can be received as a classifier by a hardware unit having a processing function. The target image input by the image classification device outputs the determined target partition model to the image classification device, which is not limited herein.

The image processing device may also generate the classifier by itself according to the training operation of the neural network. Specifically, the image processing apparatus may acquire P training images in advance, and then determine a first number of regions in each training image whose edge detection result is greater than the first threshold, and then determine, according to the first number, each training image a partitioning model; or, determining a second number of regions in each training image with a confidence greater than a second threshold, and then determining a partition model corresponding to each training image based on the second number; or, for each training session The images are partitioned using the N partition models, and the ratio of the accuracy and time of each partition model to each training image is determined, or the confidence of each partition model for partitioning each training image, or each The partition model performs the edge detection result of each training image. Then, according to the ratio, the confidence or the edge detection result, the partition model corresponding to each training image is determined. For example, the partition model with the highest ratio of precision and time for partitioning the i-th training image is the partition corresponding to the i-th image. Model, 1 ≤ i ≤ P. After determining the partition model corresponding to each training image, the BP algorithm or other algorithms can be used to obtain the classifier. As long as the P training images cover various types of images, and the P value is sufficiently large, the classifier can be guaranteed to have a classification effect that satisfies the actual needs. The classifier obtained by such a method can determine, as the target partition model of the target image, the partition model in which the ratio of the accuracy of the partitioning of the target image and the time is the largest among the N partition models.

203. Partition the target image by using a target partition model to obtain a partition result of the target image.

After determining the target partition model, the image processing apparatus partitions the target image using the target partition model to obtain a partition result of the target image.

As mentioned in the above discussion, the higher the accuracy of the model, the longer the model will perform the partitioning operation. Therefore, if the final target partition model is a model with higher precision, it will take a longer time for the image partitioning operation. In order to shorten the time of image processing, optionally, the image processing apparatus may perform step 204 while performing step 202:

204. Partition the target image using the first partition model.

The image processing apparatus partitions the target image using the first partition model, wherein the first partition model may be the most accurate model among the N partition models. It is also possible to model the longest time for image partitioning in the N partition models. It may also be a model determined in the N partition models according to other standards, which is not limited herein.

Since the accuracy of the first partition model is high and the image partitioning operation time is long, after the image processing apparatus performs the step 202 to determine the target partition model, the step 204 is not performed. At this time, the image processing apparatus performs step 203. Specifically, if the target partition model is the first partition model, the target image is further partitioned by using the first partition model, and the partition result of the first partition model is used as the partition result of the target image. In this case, since step 202 is performed simultaneously with the operation of partitioning using the first partition model, the time taken to perform step 202 can be saved. If the target partition model is not the first partition model, the image processing apparatus stops the operation of partitioning the target image using the first partition model, and partitions the target image using the determined target partition model to obtain a partition result of the target image.

As mentioned in the above discussion, different partition models use different algorithms to handle different types of images. Therefore, in practice, it is not always possible to strictly determine which model has the highest accuracy or the partition operation takes the longest. For example, suppose that for Image 1, the partitioning result obtained by using Model 1 for image partitioning has the highest accuracy; but for Image 2, the accuracy of the partitioning result obtained by using Model 1 for image partitioning may not be very high, but instead The partitioning result obtained by model 2 for image partitioning has the highest precision. Therefore, the first partition model may actually be a partition model that is manually specified in the N partition models or specified by the image processing apparatus by default. For example, the image processing device may preset the accuracy relationship of the N partition models, and the first partition model is the fine in the N partition models. The highest partition model. Alternatively, the image processing apparatus may preset a time-consuming relationship between the N partition models for performing an image partitioning operation, and the first partition model is a partition model that takes the longest image partitioning operation in the N partition models. The accuracy relationship or the time-consuming relationship may be obtained by the image processing device, or determined by the image processing device, or manually set.

Optionally, the image processing apparatus performs the step 202, in addition to using the first partition model to partition the target image in parallel, and simultaneously performing the pair using the second partition model, the third partition model, or more partition models. The operation of partitioning the target image is similar to the step 204, and is not described here. However, the more partition models that perform the partitioning operation on the image in parallel, the higher the utilization rate of the processor of the image processing apparatus, and the larger the memory usage. Therefore, in practical applications, the performance of the image processing apparatus should be integrated, and a partition model for performing image partitioning operations on the target image in parallel should be selected.

In the image processing method provided by the embodiment of the present invention, the image processing apparatus acquires a target image, determines a target partition model corresponding to the target image in the N partition models by using a classifier, and then partitions the target image by using the target partition model to obtain a target image. Partition results. The image processing apparatus in the embodiment of the present invention does not adopt a single partition model, but uses a classifier to select an appropriate partition model in the N partition models, thereby ensuring that the image processing apparatus can select accuracy for all types of images. Partition models with better speeds are used to partition the target image to meet actual needs.

After the image partitioning device partitions the target image, an image sorting operation is also required. Specifically, the image processing apparatus may classify the target image by using a fixed-size convolution kernel (English: convolution matrix or convolution kernel) by using a classification model, using a floating point or a fixed-point channel. Similar to the image partitioning operation, it is suitable to use a classification model with higher precision and slower speed for more complex images. A simpler image is suitable for a classification model with lower precision and faster speed. The complexity of the image can be reflected by the accuracy of the target partition model. Therefore, in the present invention, the image processing apparatus may determine the target size of the convolution kernel of the corresponding classification model according to the accuracy of the target partition model, and then use the convolution kernel of the target size to partition the target image. The results are classified to obtain the classification result of the target image. Such a method can classify more complex images by a larger convolution kernel, and a simpler image is classified by a smaller convolution kernel, taking into account the accuracy and speed of the classification calculation.

Alternatively, the image processing apparatus may determine the corresponding image according to the accuracy of the target partition model. The calculation channel of the classification model, for example, if the target partition model is the most accurate partition model among the N partition models (or one of the top n high-resolution partition models, n<N), the image processing apparatus can use the floating in parallel The point channel and the fixed point channel classify the segmentation result of the target image to obtain the classification result of the target image. Parallel use of floating point channels and fixed point channels can speed up the classification operation and improve the efficiency of image classification. If the target partition model is not the most accurate model of the N partition models (or one of the top n high-precision partition models), the image processing apparatus may classify the partition results of the target image using only the floating point channel.

The embodiment shown in FIG. 2 introduces an image processing method provided by the present invention. An image processing apparatus for implementing the above method will be described below. The basic structure of the image processing apparatus is as follows:

The image acquisition module 301 is configured to perform step 201 in the embodiment shown in FIG. 2;

The model determining module 302 is configured to perform step 202 in the embodiment shown in FIG. 2;

The image recognition module 303 is configured to perform step 203 in the embodiment shown in FIG. 2 . Optionally, step 204 can also be performed.

For the functions of the modules of the image processing apparatus, reference may be made to the description in the method embodiment shown in FIG. 2, and details are not described herein.

Optionally, the image recognition module 303 is further configured to determine a target size of the convolution kernel corresponding to the target partition model according to the precision of the target partition model, and use the convolution kernel of the target size to classify the partition result of the target image. Get the classification results.

Optionally, the image recognition module is further configured to perform classification of the target image by using a floating point channel and a fixed point channel in parallel when the target partition model is not the highest precision partition model of the N partition models. .

The embodiment of the present invention further provides a computing device 400 for implementing the image processing method in the embodiment shown in FIG. 2. See Figure 4 for the basic structure. The computing device specifically includes a processor 401, a memory 402, a bus 403, and a communication interface 404. The processor 401, the memory 402, and the communication interface 404 can implement communication connection with each other through the bus 403, and can also implement communication by other means such as wireless transmission.

The memory 402 memory may include a volatile memory (English: volatile memory), such as random access memory (English: random-access memory, abbreviation: RAM); the memory may also include non-volatile memory (English: non-volatile memory) ), such as read-only memory (English: Read-only memory, abbreviated as: ROM), flash memory (English: flash memory), hard disk (English: hard disk drive, abbreviation: HDD) or solid state drive (English: solid-state drive, abbreviation: SSD); memory 402 Combinations of the above types of memory may also be included. When the computing device 400 is running, the memory 402 loads the N partition models, the accuracy relationship or the time-consuming relationship of the N partition models, the classifier, each classification model, and the like for use by the processor 401. When the technical solution provided by the present invention is implemented by software, the program code for implementing the image processing method provided by the present invention may be stored in the memory 402 and executed by the processor 401.

Computing device 400 acquires the target image via communication interface 404 and returns the classification result for the target image to the user via communication interface 404.

The processor 401 can be a central processing unit (English: central processing unit, CPU for short), a graphics processing unit (English: graphics processing unit (abbreviation: GPU), digital signal processing (English: digital signal processing, abbreviation: DSP), A combination of any one or more of the hardware units having processing functions, such as a field-programmable gate array (English: field-programmable gate array). The processor 401 is mainly configured to acquire a target image to be processed; use a classifier to determine a target partition model corresponding to the target image, and use the first partition model to partition the target image; after determining the target partition model, use the target partition model pair The target image is partitioned to obtain a partition result of the target image; according to the accuracy of the target partition model, the target size of the convolution kernel of the corresponding classification model and the corresponding floating point and/or fixed point calculation channel are determined, and then the target size is used. The convolution kernel and the calculation channel classify the partition result of the target image to obtain the classification result of the target image.

In the several embodiments provided by the present application, it should be understood that the disclosed computing device, apparatus, and method may be implemented in other manners. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there may be another division manner, for example, multiple modules or components may be combined or Can be integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interface, device or unit, and may be in an electrical, mechanical or other form.

In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each module may exist physically separately, or two or more modules may be integrated in one In the unit. The above integrated modules can be implemented in the form of hardware or in the form of software functional units.

The integrated modules, if implemented in the form of software functional units and sold or used as separate products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention, which is essential or contributes to the prior art, or all or part of the technical solution, may be embodied in the form of a software product stored in a storage medium. A number of instructions are included to cause a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.

The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to be limiting; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art will understand that The technical solutions described in the embodiments are modified, or the equivalents of the technical features are replaced by the equivalents of the technical solutions of the embodiments of the present invention.

Claims

An image processing method, which is applicable to an image processing apparatus, and includes:

Obtaining a target image to be processed;

In the N partition models, a classifier is used to determine a target partition model corresponding to the target image, and the N is an integer greater than 1;

The target image is partitioned using the target partition model to obtain a partition result of the target image.
The image processing method according to claim 1, wherein the method further comprises: when the using the classifier to determine a target partition model corresponding to the target image,

Partitioning the target image using a first partition model, wherein the first partition model is the highest-precision partition model among the N partition models or a partition model that takes the longest time to perform an image partitioning operation;

The partitioning the target image by using the target partition model, and obtaining the partition result of the target image includes:

If the target partition model is the first partition model, continue to partition the target image by using the first partition model, and use a partition result of the first partition model as a partition result of the target image. ;

If the target partition model is not the first partition model, stopping the operation of partitioning the target image by using a first partition model, and partitioning the target image by using the target partition model to obtain The partition result of the target image.
The image processing method according to claim 1 or 2, wherein the method further comprises: before the acquiring the target image to be processed:

The classifier is obtained by an image training operation of a neural network, and the classifier is configured to determine a target partition model corresponding to the target image in the N partition models.
The image processing method according to claim 3, wherein the obtaining the classifier by the image training operation of the neural network comprises:

Obtain P training images;

And dividing the P training images by using the N partition models, and respectively determining a ratio of accuracy and time of each partition image to partitioning each training image;

Determining a partition model corresponding to each training image according to a ratio of accuracy and time of each training image to each of the training models, wherein the partition model corresponding to the i-th training image in the P training images a partitioning model having the largest ratio of precision to time for partitioning the i-th image, 1 ≤ i ≤ P;

Performing training according to the partition model corresponding to the P training images, and obtaining the classifier, wherein the classifier is configured to maximize a ratio of accuracy and time of partitioning the target image in the N partition models. The partition model is determined as a target partition model corresponding to the target image.
The image processing method according to any one of claims 1 to 4, further comprising:

Determining, according to an accuracy of the target partition model, a target size of a convolution kernel corresponding to the target partition model;

Using the convolution kernel of the target size, classifying the partitioning result of the target image to obtain a classification result of the target image.
The image processing method according to any one of claims 1 to 5, further comprising:

If the target partition model is not the most accurate partition model among the N partition models, the partitioning results of the target image are classified in parallel using a floating point channel and a fixed point channel to obtain a classification result of the target image.
An image processing apparatus, comprising:

An image acquisition module, configured to acquire a target image to be processed;

a model determining module, configured to determine, in the N partition models, a target partition model corresponding to the target image by using a classifier, where N is an integer greater than 1;

And an image recognition module, configured to partition the target image by using the target partition model to obtain a partition result of the target image.
The image processing device according to claim 1, wherein the image recognition module is specifically configured to:

And when the using the classifier to determine a target partition model corresponding to the target image, partitioning the target image by using a first partition model, wherein the first partition model is the highest precision among the N partition models The partition model or the partition model that takes the longest time for image partitioning;

After the model determining module obtains the partitioning result of the target image, if the target partitioning model is the first partitioning model, continuing to partition the target image by using the first partitioning model, and The partition result of the first partition model is used as a partition result of the target image; if the target partition model is not the first partition model, stopping the partitioning of the target image using the first partition model And partitioning the target image using the target partition model to obtain a partition result of the target image.
The image processing apparatus according to claim 7 or 8, wherein the model determining module is further configured to:

The classifier is obtained by an image training operation of a neural network, and the classifier is configured to determine a target partition model corresponding to the target image in the N partition models.
The image processing device according to claim 9, wherein the model determining module is further configured to:

Obtain P training images;

And dividing the P training images by using the N partition models, and respectively determining a ratio of accuracy and time of each partition image to partitioning each training image;

Determining a partition model corresponding to each training image according to the ratio of the accuracy of the partitioning of each training image to each of the partition models, wherein the partition model corresponding to the i-th training image is: for the ith sheet The partition model with the largest ratio of accuracy and time of image partitioning, 1≤i≤P;

Performing training according to the partition model corresponding to the P training images, and obtaining the classifier, wherein the classifier is configured to maximize a ratio of accuracy and time of partitioning the target image in the N partition models. The partition model is determined as a target partition model corresponding to the target image.
The image processing device according to any one of claims 7 to 10, wherein the image recognition module is further configured to:

Determining, according to an accuracy of the target partition model, a target size of a convolution kernel corresponding to the target partition model; using a convolution kernel of the target size, classifying a partition result of the target image to obtain the target image Classification results.
The image processing device according to any one of claims 7 to 11, wherein the image recognition module is further configured to:

When the target partition model is not the highest precision partition model among the N partition models, The partitioning results of the target image are classified in parallel using a floating point channel and a fixed point channel to obtain a classification result of the target image.
A computer device, comprising: a processor, a memory, and a communication interface, by calling an instruction in the memory, the processor is configured to:

Obtaining a target image to be processed;

In the N partition models, a classifier is used to determine a target partition model corresponding to the target image, and the N is an integer greater than 1;

The target image is partitioned using the target partition model to obtain a partition result of the target image.
The computing device of claim 13, wherein the processor is further configured to:

Partitioning the target image using a first partition model, wherein the first partition model is the highest-precision partition model among the N partition models or a partition model that takes the longest time to perform an image partitioning operation;

The partitioning the target image by using the target partition model, and obtaining the partition result of the target image includes:

If the target partition model is the first partition model, continue to partition the target image by using the first partition model, and use a partition result of the first partition model as a partition result of the target image. ;

If the target partition model is not the first partition model, stopping the operation of partitioning the target image by using a first partition model, and partitioning the target image by using the target partition model to obtain The partition result of the target image.
The computing device of claim 13 or 14, wherein the processor is further configured to:

The classifier is obtained by an image training operation of a neural network, and the classifier is configured to determine a target partition model corresponding to the target image in the N partition models.
The computing device of claim 15, wherein the processor is further configured to:

Obtain P training images;

And dividing the P training images by using the N partition models, and respectively determining a ratio of accuracy and time of each partition image to partitioning each training image;

Determining a partition model corresponding to each training image according to the ratio of the accuracy of the partitioning of each training image to each of the partition models, wherein the partition model corresponding to the i-th training image is: for the ith sheet The partition model with the largest ratio of accuracy and time of image partitioning, 1≤i≤P;

Performing training according to the partition model corresponding to the P training images, and obtaining the classifier, wherein the classifier is configured to maximize a ratio of accuracy and time of partitioning the target image in the N partition models. The partition model is determined as a target partition model corresponding to the target image.
The computing device according to any one of claims 13 to 16, wherein the processor is further configured to:

Determining, according to an accuracy of the target partition model, a target size of a convolution kernel corresponding to the target partition model;

Using the convolution kernel of the target size, classifying the partitioning result of the target image to obtain a classification result of the target image.
The computing device according to any one of claims 13 to 17, wherein the processor is further configured to:

If the target partition model is not the most accurate partition model among the N partition models, the partitioning results of the target image are classified in parallel using a floating point channel and a fixed point channel to obtain a classification result of the target image.