WO2023185497A1 - 组织图像的识别方法、装置、可读介质和电子设备 - Google Patents

组织图像的识别方法、装置、可读介质和电子设备 Download PDF

Info

Publication number
WO2023185497A1
WO2023185497A1 PCT/CN2023/082017 CN2023082017W WO2023185497A1 WO 2023185497 A1 WO2023185497 A1 WO 2023185497A1 CN 2023082017 W CN2023082017 W CN 2023082017W WO 2023185497 A1 WO2023185497 A1 WO 2023185497A1
Authority
WO
WIPO (PCT)
Prior art keywords
sample
camouflage
image
salient
confidence
Prior art date
Application number
PCT/CN2023/082017
Other languages
English (en)
French (fr)
Inventor
边成
李永会
张志诚
Original Assignee
北京字节跳动网络技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京字节跳动网络技术有限公司 filed Critical 北京字节跳动网络技术有限公司
Publication of WO2023185497A1 publication Critical patent/WO2023185497A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Definitions

  • the present disclosure relates to the field of image processing technology, and specifically, to a tissue image recognition method, device, readable medium, electronic equipment, computer program product, and computer program.
  • Endoscopy as a common examination method in the medical field, has been widely used because it can directly observe the internal conditions of the human body.
  • tissue images inside the human body can be collected in real time.
  • image recognition the areas where polyps, ulcers and other objects are located can be identified from the tissue images.
  • image recognition mainly targets late-stage polyps. Late-stage polyps are larger and easier to find, so the accuracy of identifying polyps is high.
  • the accuracy of identifying polyps is low and cannot meet the actual needs of endoscopy.
  • the present disclosure provides a method for identifying tissue images, the method including:
  • the tissue image is identified according to the salient image features, the camouflage image features, the salience confidence level and the camouflage confidence level to determine a target recognition result, and the target recognition result is used to identify the tissue The area in the image where the target object is located.
  • the present disclosure provides a device for identifying tissue images, the device including:
  • An acquisition module is used to acquire tissue images collected by the endoscope
  • An extraction module used to respectively extract significant image features and camouflage image features of the tissue image
  • a confidence determination module configured to determine the significant confidence of the tissue image based on the salient image features, and determine the camouflage confidence of the tissue image based on the camouflage image features;
  • An identification module configured to identify the tissue image according to the salient image features, the camouflage image features, the salient confidence level and the camouflage confidence level to determine a target recognition result, where the target recognition result is To identify the area where the target object is located in the tissue image.
  • the present disclosure provides a computer-readable medium having a computer program stored thereon, and when the program is executed by a processing device, the steps of the method described in the first aspect of the present disclosure are implemented.
  • an electronic device including:
  • a processing device configured to execute the computer program in the storage device to implement the steps of the method described in the first aspect of the present disclosure.
  • the present disclosure provides a computer program product, including a computer program that implements the steps of the method described in the first aspect of the present disclosure when executed by a processing device.
  • the disclosure provides a computer program that, when executed by a processing device, implements the steps of the method described in the first aspect of the disclosure.
  • the present disclosure first obtains the tissue image collected by the endoscope, then extracts the significant image features and camouflage image features of the tissue image, and then determines the significant confidence of the tissue image based on the salient image features, and determines the tissue based on the camouflage image features.
  • the camouflage confidence of the image is recognized by combining the salient image features, camouflage image features, salience confidence and camouflage confidence to obtain the area where the target object is located in the tissue image.
  • This disclosure respectively extracts the salient image features and camouflaged image features of the tissue image, and further determines the salient confidence and camouflage confidence, thereby determining the area where the target object is located in the tissue image, which can improve the accuracy and generalization ability of image recognition.
  • Figure 1 is a flow chart of a method for identifying tissue images according to an exemplary embodiment
  • Figure 2 is a schematic diagram of an identification network according to an exemplary embodiment
  • Figure 3 is a flow chart of another method for identifying tissue images according to an exemplary embodiment
  • Figure 4 is a flow chart of another method for identifying tissue images according to an exemplary embodiment
  • Figure 5 is a schematic diagram of a saliency encoder, a disguised encoder, a saliency classifier, a disguised classifier, a decoder and a discriminator according to an exemplary embodiment
  • Figure 6 is a flow chart for jointly training a saliency encoder, a camouflage encoder, a saliency classifier, a camouflage classifier and a decoder according to an exemplary embodiment
  • Figure 7 is a flow chart illustrating another joint training of a saliency encoder, a camouflage encoder, a saliency classifier, a camouflage classifier and a decoder according to an exemplary embodiment
  • FIG. 8 is a block diagram of a tissue image recognition device according to an exemplary embodiment
  • FIG. 9 is a block diagram of another tissue image recognition device according to an exemplary embodiment.
  • Figure 10 is a block diagram of another device for identifying tissue images according to an exemplary embodiment
  • FIG. 11 is a block diagram of an electronic device according to an exemplary embodiment.
  • the term “include” and its variations are open-ended, ie, “including but not limited to.”
  • the term “based on” means “based at least in part on.”
  • the term “one embodiment” means “at least one embodiment”; the term “another embodiment” means “at least one additional embodiment”; and the term “some embodiments” means “at least some embodiments”. Relevant definitions of other terms will be given in the description below.
  • Figure 1 is a flow chart of a method for identifying tissue images according to an exemplary embodiment. As shown in Figure 1, the method includes the following steps:
  • Step 101 Obtain tissue images collected by the endoscope.
  • the endoscope will continuously collect images of human tissue according to a preset collection cycle, that is, tissue images in this embodiment.
  • the tissue image may be an image collected by an endoscope at the current moment or an image collected by an endoscope at any time. That is to say, the tissue image may be an image collected during the insertion process of the endoscope, or may be an image collected during the withdrawal process of the endoscope, which is not specifically limited in this disclosure.
  • the endoscope described in the embodiments of the present disclosure may be, for example, an enteroscope, a colonoscope, a gastroscope, etc. If the endoscope is a enteroscope, then the above tissue image is an intestinal image.
  • the endoscope can also be used to collect images of other tissues, which is not specifically limited in this disclosure.
  • tissue image After obtaining the tissue image, you can first determine whether the tissue image is valid to filter out invalid tissue images. If the tissue image is an invalid image, the tissue image can be directly discarded and the tissue image collected in the next acquisition cycle can be continued to be obtained. This can reduce unnecessary data processing and increase processing speed. For example, pre-trained filtering models can be utilized to filter tissue images to remove invalid tissue images.
  • the structure of the filtering model can be, for example, CNN (English: Convolutional Neural Networks, Chinese: Convolutional Neural Network) or LSTM (English: Long Short-Term Memory, Chinese: Long Short-Term Memory Network), or it can be a Transformer (such as Vision Transformer) Encoder, etc., this disclosure does not specifically limit this.
  • Step 102 Extract significant image features and camouflage image features of the tissue image respectively.
  • the target object can be understood as the object that needs attention during the endoscopy process, and can be understood as the foreign body growing on the tissue, such as polyps, ulcers, etc.
  • the growth process of target objects is often from small to Large, correspondingly, the difficulty of being recognized increases from large to small. If the target object can be found at an early stage, the target object can be resected or treated in time to avoid further deterioration of the target object.
  • polyps in the colon As an example, the development process of polyps is often a process of gradual enlargement from normal mucosa - mucosal epithelial hyperplasia - adenomatous polyps - adenomatous polyps.
  • the characteristics of the target object in the early stage often have camouflage characteristics and are not easy to be recognized, that is, the target object is a camouflage type target object in the early stage.
  • the characteristics of the target object in the later stage are often very significant and easy to be recognized, that is, the target object belongs to a salient type of target object in the later stage. Therefore, two different feature extraction methods can be used to extract significant image features and camouflage image features of tissue images.
  • the method of extracting salient image features can be understood as a feature extraction method suitable for situations where the tissue image includes a salient type of target object, and the extracted salient image features can reflect the characteristics of the salient type of target object.
  • the method of extracting camouflage image features can be understood as a feature extraction method suitable for situations where the tissue image includes a camouflage type target object, and the extracted camouflage image features can reflect the characteristics of the camouflage type target object.
  • the saliency extraction model (including the saliency encoder mentioned later) can be trained based on the sample image including the target object of the salient type, and is used to extract the salient image features, and based on the sample image including the target object of the camouflage type, Train the camouflage extraction model (including the camouflage encoder mentioned later) to extract camouflage image features.
  • the structure of the salient extraction model and the camouflage extraction model can be: CNN or VGG (English: Visual Geometry Group) network, or it can be Encoder, ResNet, DenseNet, etc. in Transformer. This disclosure does not specifically limit this.
  • Step 103 Determine the salience confidence of the tissue image based on the salient image features, and determine the camouflage confidence of the tissue image based on the camouflage image features.
  • the target object in the tissue image is a salient type or a camouflaged type based on the salient image features and camouflaged image features. Since the saliency of the target object changes linearly rather than binary, the saliency (or camouflage degree) of the target object in the tissue image can be expressed by confidence. Specifically, the salience confidence of the tissue image can be determined based on the salient image features. The salience confidence is used to characterize the probability that the target object in the tissue image is a salient type, and can also be understood as the salience degree of the target object in the tissue image.
  • the camouflage confidence of the tissue image can be determined based on the characteristics of the camouflage image.
  • the camouflage confidence is used to characterize the probability that the target object in the tissue image is a camouflage type, and can also be understood as the degree of camouflage of the target object in the tissue image.
  • the saliency prediction model (including the saliency classifier mentioned later) can be trained based on the sample image including the target object of the salient type to determine the saliency confidence, and based on the sample image including the target object of the camouflage type, Train a camouflage prediction model to determine camouflage confidence.
  • the structure of the saliency prediction model and the camouflage prediction model can be, for example, a SoftMax model, a decision tree model, or an SVM (English: Support Vector Machine, Chinese: Support Vector Machine), etc. This disclosure does not specifically limit this.
  • Step 104 Recognize the tissue image according to the salient image features, camouflage image features, salience confidence and camouflage confidence to determine the target recognition result.
  • the target recognition result is used to identify the area where the target object is located in the tissue image.
  • the tissue image can be recognized by combining salient image features, camouflage image features, salience confidence and camouflage confidence to obtain a target recognition result, where the target recognition result is used to indicate which areas in the tissue image belong to the target object.
  • the recognition process can be understood as the process of annotating tissue images, marking out the pixels belonging to the target object in the tissue image.
  • the recognition process can also be understood as the process of segmenting the tissue image, extracting the pixels belonging to the target object in the tissue image.
  • the target recognition result can be a binary image with the same size as the tissue image, in which the white area corresponds to the target object (or the black area corresponds to the target object).
  • the target recognition result can also be a set of coordinate ranges used to represent the tissue image. The coordinate range of the target object.
  • the salient confidence and the camouflage confidence can be compared to determine whether to identify based on salient image features or camouflaged image features. For example, if the salience confidence is high and the camouflage confidence is low, it means that the probability of the target object in the tissue image being a salient type is high, and then the target object can be identified based on the salient image features to obtain the target recognition result. Ruoxian If the confidence level is low and the camouflage confidence level is high, it means that the probability of the target object in the tissue image being a camouflage type is high. Then it can be identified based on the characteristics of the camouflage image to obtain the target recognition result.
  • Salient image features and camouflage image features can also be fused using salience confidence and camouflage confidence, and then recognition is performed based on the fused image features to obtain target recognition results. It is also possible to identify based on salient image features and identify based on camouflaged image features, and then use the salient confidence and camouflage confidence to fuse the two recognition results to obtain the target recognition result. In this way, the target recognition result takes into account both salient image features and camouflaged image features, which can effectively improve the recognition accuracy of camouflaged target objects, thereby improving the generalization ability of image recognition.
  • the pre-jointly trained saliency encoder (denoted as E s ), camouflage encoder (denoted as E c ), saliency classifier (denoted as C s ), and It is implemented by pretending to be a classifier (denoted as C c ) and a decoder (denoted as Decoder).
  • the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier and decoder can be understood as a recognition network used to recognize tissue images.
  • the structural diagram of the recognition network is shown in Figure 2. Among them, the recognition network is jointly trained based on a preset sample image set.
  • the sample image set includes multiple salient sample images and multiple camouflage sample images.
  • the target object in the salient sample image is a salient type
  • the target object in the camouflage sample image is a camouflage type. .
  • step 102 can be implemented as:
  • the pre-trained saliency encoder is used to extract features from the tissue image to obtain salient image features
  • the pre-trained camouflage encoder is used to extract features from the tissue image to obtain camouflage image features.
  • the tissue image can be used as the input of the saliency encoder, and the output of the saliency encoder is the salient image feature.
  • the tissue image can be used as the input of the camouflage encoder, and the data of the camouflage encoder is the camouflage image feature.
  • the salient encoder and the camouflage encoder are both used to extract features from the image, and the ResNet50 structure can be used.
  • Step 103 can be implemented as follows:
  • the salient image features can be input into the salient classifier, and the output of the salient classifier is the salient confidence (expressed as salient confidence).
  • the camouflage image features can be input into the camouflage classifier, and the output of the camouflage classifier is the camouflage confidence. Degree (expressed as camouflaged confidence).
  • the saliency classifier is used to determine the matching probability between the tissue image and the saliency type, that is, the saliency confidence.
  • the camouflage classifier is used to determine the probability of matching the tissue image to the camouflage type, that is, the camouflage confidence.
  • Step 104 can be implemented as follows:
  • the pre-trained decoder is used to identify tissue images based on salient image features, camouflage image features, salience confidence, and camouflage confidence to determine the target recognition results.
  • the decoder can be used to identify tissue images by combining salient image features, camouflaged image features, salient confidence, and camouflage confidence to obtain target recognition results.
  • the implementation of using the decoder to identify tissue images is described in detail below.
  • FIG. 3 is a flow chart of another method for identifying tissue images according to an exemplary embodiment. As shown in Figure 3, the implementation of step 104 may include:
  • Step 1041 Recognize the tissue image according to salient image features to determine the salient recognition result.
  • Step 1042 Recognize the tissue image according to the characteristics of the camouflage image to determine the camouflage recognition result.
  • Step 1043 Fusion of the salient recognition results and the camouflage recognition results based on the saliency confidence and camouflage confidence to determine the target recognition result.
  • the salient recognition results and camouflage recognition results output by the decoder can be fused according to the saliency confidence and camouflage confidence, and the fusion result can be used as the target recognition result.
  • the significant confidence level can be used as the weight of the significant recognition result
  • the camouflage confidence level can be used as the weight of the camouflage recognition result, so as to perform a weighted sum of the salient recognition results and the camouflage recognition results, and the result obtained is the target recognition result.
  • FIG. 4 is a flow chart of another tissue image recognition method according to an exemplary embodiment. As shown in Figure 4, the implementation of step 104 may also include:
  • Step 1044 If the significant confidence is greater than the first threshold and the camouflage confidence is less than or equal to the second threshold, identify the tissue image according to the salient image features to determine the target recognition result.
  • Step 1045 If the camouflage confidence is greater than the second threshold and the significant confidence is less than or equal to the first threshold, identify the tissue image according to the camouflage image features to determine the target recognition result.
  • Step 1046 if the significant confidence is less than or equal to the first threshold, and the camouflage confidence is less than or equal to the second threshold, identify the tissue image based on the salient image features to determine the significant recognition result; identify the tissue image based on the camouflaged image features , to determine the camouflage recognition result; according to the salience confidence and camouflage confidence, the salient recognition results and the camouflage recognition results are fused to determine the target recognition result.
  • first threshold For example, you can first compare the significant confidence with a preset first threshold, and at the same time compare the camouflage confidence with a preset second threshold, and then determine how to identify the tissue image based on the comparison results.
  • the first threshold and the second The thresholds can be the same or different.
  • the first threshold may be 0.7
  • the second threshold may be 0.6.
  • the salience confidence is greater than the first threshold and the camouflage confidence is less than or equal to the second threshold, it means that the probability of the target object in the tissue image being a salient type is high, then the salient image features can be input into the decoder to achieve Organization image recognition, at this time the output of the decoder is the target recognition result.
  • the camouflage confidence is greater than the second threshold and the significant confidence is less than or equal to the first threshold, it means that the probability of the target object in the tissue image being a camouflage type is high, then the camouflage image features can be input into the decoder to achieve Organization image recognition, at this time the output of the decoder is the target recognition result.
  • the salient image features can be input Decoder, at this time the decoder outputs significant recognition results.
  • the camouflage image features are then input into the decoder, and the decoder outputs the camouflage recognition result.
  • the salient recognition results and camouflage recognition results output by the decoder can be fused according to the saliency confidence and camouflage confidence, and the fusion result can be used as the target recognition result.
  • the significant confidence level can be used as the weight of the significant recognition result
  • the camouflage confidence level can be used as the weight of the camouflage recognition result, so as to perform a weighted sum of the salient recognition results and the camouflage recognition results, and the result obtained is the target recognition result.
  • the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier and decoder When jointly training the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier and decoder, the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier can be trained together with the preset discriminator
  • the connection relationship between the decoder, the decoder and the discriminator is shown in Figure 5.
  • Figure 6 is a flow chart for jointly training a saliency encoder, a camouflage encoder, a saliency classifier, a camouflage classifier and a decoder according to an exemplary embodiment.
  • the saliency encoder, camouflage encoder , the saliency classifier, the camouflage classifier and the decoder are jointly trained in the following way:
  • Step A Obtain a sample input set and a sample output set.
  • the sample input set includes: multiple sample inputs.
  • Each sample input includes a sample image in the sample image set.
  • the sample image is a salient sample image or a camouflage sample image.
  • the sample output set includes sample output corresponding to each sample input.
  • the sample output is the real recognition result of the corresponding sample image. The real recognition result is used to identify the area where the target object is located in the corresponding sample image.
  • the sample input set includes: multiple sample inputs.
  • Each sample input includes a sample image in the sample image set.
  • the sample image can be a salient sample image (expressed as salient data) or a camouflaged sample image (expressed as camouflaged data). ).
  • the sample output set includes sample output corresponding to each sample input.
  • the sample output is the real recognition result of the corresponding sample image.
  • the real recognition result is used to represent the area where the target object is located in the corresponding sample image.
  • the real recognition result can be binary An image, which can also be a coordinate range.
  • the sample image may be a large number of tissue images acquired during a previous endoscopic examination.
  • the sample input set and the sample output set can be divided into a training set, a verification set, and a test set in proportions of 60%, 10%, and 30%. This disclosure does not specifically limit the proportional distribution.
  • Step B input each sample as the input of the salience encoder to obtain the salient sample image features, and input the sample as the input of the camouflage encoder to obtain the camouflage sample image features.
  • Step C Input the salient sample image features into the saliency classifier to obtain the first confidence level, and input the camouflage sample image features into the camouflage classifier to obtain the second confidence level.
  • the sample input is used as the input of the salient encoder and the camouflage encoder respectively, and the salient sample image features output by the saliency encoder and the camouflage sample image features output by the camouflage encoder are obtained.
  • the salient sample image features can be input into the salient classifier to obtain the first confidence level output by the salient classifier
  • the camouflage sample image features can be input into the camouflage classifier to obtain the second confidence level output by the camouflage classifier.
  • the first confidence level is used to characterize the probability that the target object in the sample image is a salient type, and can also be understood as the salience degree of the target object in the sample image.
  • the second confidence level is used to characterize the probability that the target object in the sample image is a camouflage type, and can also be understood as the degree of camouflage of the target object in the sample image.
  • Step D input the target sample image features into the decoder to obtain the sample recognition result output by the decoder. If the sample input is a significant sample image, the target sample image feature is a significant sample image feature. If the sample input is a camouflage sample image, The target sample image features are the camouflage sample image features.
  • Step E Use the sample recognition result and the sample output corresponding to the sample input as the input of the discriminator to obtain the discrimination result output by the discriminator.
  • the target sample image features are determined, and then the target sample image features are input to the decoder to obtain the sample recognition result output by the decoder, and the sample recognition result is used to represent the decoder Predict the area where the target object is located in the sample image, which can be a binary image or a coordinate range.
  • the target sample image feature is a salient sample image feature
  • the target sample image feature is a camouflage sample image feature.
  • the sample recognition result and the sample output corresponding to the sample input can be used as the input of the discriminator (denoted as Discriminator) to obtain the discrimination result output by the discriminator.
  • the discriminant result is used to indicate that the input is a salient type of uncertainty (expressed as salient uncertainty), or a disguised type of uncertainty (expressed as camouflaged uncertainty).
  • the sample recognition result is input into the discriminator, the discriminator outputs the first discrimination result, the sample is input into the corresponding sample output and input into the discriminator, and the discriminator outputs the second discrimination result.
  • the first discrimination result is made as accurate as possible It is different from the second discrimination result to achieve the purpose of confrontation training.
  • the discriminator can be the discriminator in GAN (English: Generative Adversarial Networks, Chinese: Generative Adversarial Networks), and the structure can be ResNet34.
  • Step F Determine the total loss based on the significant sample image features, camouflaged sample image features, first confidence level, second confidence level, sample recognition results, and discrimination results.
  • Step G jointly train the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier, decoder and discriminator according to the total loss.
  • the total loss can be determined based on the salient sample image features, the disguised sample image features, the first confidence level, the second confidence level, the sample recognition results, and the discrimination results. Then, with the goal of reducing the total loss, the backpropagation algorithm is used to jointly train the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier, decoder and discriminator.
  • the initial learning rate for jointly training the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier, decoder and discriminator can be set to: 1e-4
  • the Batch size can be set to: 32
  • the optimizer can be selected :Adam
  • Epoch can be set to: 100
  • the size of the sample image can be: 256 ⁇ 256.
  • FIG. 7 is another flow chart for jointly training a saliency encoder, a camouflage encoder, a saliency classifier, a camouflage classifier and a decoder according to an exemplary embodiment.
  • step F can be performed in the following manner to fulfill:
  • Step F1 Determine the similarity loss based on the salient sample image features and the camouflaged sample image features.
  • the similarity loss can be determined based on the cosine similarity of the salient sample image features and the camouflaged sample image features.
  • L s represents the similarity loss
  • S s represents the cosine similarity between the salient sample image features and the camouflage sample image features when the sample input is a salient sample image
  • S c represents the cosine similarity between the salient sample image features and the camouflage sample image features when the sample input is a camouflage sample image,
  • represents the norm.
  • Step F2 Determine the confidence loss based on the first confidence level and the second confidence level.
  • the cross-entropy loss of the first confidence level and the first target true value can be summed with the cross-entropy loss of the second confidence level and the second target true value to obtain the confidence loss, the first target true value and the second target true value.
  • the true value of the second target is determined based on the corresponding sample input.
  • L c represents the confidence loss
  • f ce represents the cross entropy (English: Cross Entropy) loss
  • C s () represents the processing of the significant classifier
  • C c () represents the processing of the camouflage classifier
  • x s represents the significant sample image
  • x c represents the disguised sample image.
  • C s (x s ) indicates that when the sample input is a significant sample image, the output of C s is the first confidence level corresponding to the significant sample image.
  • the first target true value is 1, that is to say, when the sample input is a significant sample image sample image, the output of C s is expected to be 1.
  • C s (x c ) indicates that when the sample input is a disguised sample image, the output of C s is the first confidence corresponding to the disguised sample image.
  • the first target true value is 0, that is to say, when the sample input is a disguised sample image sample image, the output of C s is expected to be 0.
  • C c (x s ) indicates that when the sample input is a significant sample image, the output of C c is the second confidence level corresponding to the significant sample image.
  • the second target true value is 0, that is to say, when the sample When the input is a salient sample image, the output of C c is expected to be 0.
  • C c (x c ) indicates that when the sample input is a disguised sample image, the output of C c is the second confidence corresponding to the disguised sample image.
  • the second target true value is 1, that is to say, when the sample input is a disguised sample image sample image, the output of C c is expected to be 1.
  • Step F3 Determine the recognition loss based on the sample recognition results and the sample output corresponding to each sample input.
  • the recognition loss can be obtained by summing the cross-entropy loss between the sample recognition result and the sample output corresponding to each sample input, and the Dice loss between the sample recognition result and the sample output corresponding to each sample input.
  • the recognition loss can be determined by Equation 3:
  • L p represents the recognition loss
  • f bce represents the binary cross entropy (English: Binary Cross Entropy) loss
  • f dice represents the Dice loss
  • Y s represents the sample output corresponding to the salient sample image.
  • Y c represents the sample output corresponding to the camouflage sample image.
  • Step F4 Determine the adversarial loss based on the discrimination results and the sample output corresponding to each sample input.
  • the cross-entropy loss between the discrimination result and the third target true value can be summed, and the cross-entropy loss between the sample output corresponding to each sample input and the fourth target true value can be summed to obtain the adversarial loss, the third target true value and The fourth target true value is determined based on the corresponding sample input.
  • the adversarial loss can be determined by formula 4:
  • L AD represents the adversarial loss
  • f bce represents the binary cross-entropy loss
  • D() represents the processing of the discriminator.
  • Y s represents the sample output corresponding to the significant sample image
  • the third target true value is 0. That is to say, when the sample recognition result is input to the Discriminator, the output of the Discriminator is expected to be 0 (the discrimination is false).
  • D(Y s ) represents the discrimination result corresponding to the sample output corresponding to the salient sample image.
  • the fourth target true value is 1. That is to say, when the sample output corresponding to the salient sample image is input to the Discriminator, the output of the Discriminator is expected to be 1. (judged to be true).
  • the Decoder Indicates the output of the Decoder when the sample input is a camouflage sample image, that is, the sample recognition result corresponding to the camouflage sample image.
  • Y c represents the sample output corresponding to the camouflage sample image.
  • the discrimination result corresponding to the sample recognition result corresponding to the disguised sample image At this time, the third target true value is 0. That is to say, when the sample recognition result is input to the Discriminator, the output of the Discriminator is expected to be 0 (the discrimination is false).
  • D(Y c ) represents the discrimination result corresponding to the sample output corresponding to the disguised sample image.
  • the fourth target true value is 1. That is to say, when the sample output corresponding to the disguised sample image is input to the Discriminator, the output of the Discriminator is expected to be 1. (judged to be true).
  • Step F5 Determine the total loss based on similarity loss, confidence loss, recognition loss, and confrontation loss.
  • the similarity loss, confidence loss, recognition loss, and adversarial loss can be summed to obtain the total loss, or the similarity loss, confidence loss, recognition loss, and adversarial loss can be weighted and summed to obtain the total loss.
  • L represents the total loss
  • the present disclosure first obtains the tissue images collected by endoscope, and then extracts the significant image features of the tissue images respectively. and camouflage image features, then determine the significant confidence of the tissue image based on the salient image features, determine the camouflage confidence of the tissue image based on the camouflage image features, and finally, combine the salient image features, camouflage image features, significant confidence, and camouflage confidence to Recognize the tissue image to obtain the area where the target object is located in the tissue image.
  • This disclosure respectively extracts the salient image features and camouflaged image features of the tissue image, and further determines the salient confidence and camouflage confidence, thereby determining the area where the target object is located in the tissue image, which can improve the accuracy and generalization ability of image recognition.
  • FIG. 8 is a block diagram of a tissue image recognition device according to an exemplary embodiment. As shown in Figure 8, the device 200 may include:
  • the acquisition module 201 is used to acquire tissue images collected by the endoscope.
  • the extraction module 202 is used to respectively extract salient image features and camouflage image features of the tissue image.
  • the confidence determination module 203 is configured to determine the salience confidence of the tissue image based on the salient image features, and determine the camouflage confidence of the tissue image based on the camouflage image features.
  • the identification module 204 is used to identify the tissue image according to the salient image features, camouflage image features, salience confidence and camouflage confidence to determine the target recognition result.
  • the target recognition result is used to identify the area where the target object is located in the tissue image.
  • FIG. 9 is a block diagram of another tissue image recognition device according to an exemplary embodiment.
  • the recognition module 204 may include:
  • the first identification sub-module 2041 is used to identify tissue images according to salient image features to determine salient recognition results.
  • the second identification sub-module 2042 is used to identify the tissue image according to the characteristics of the disguised image to determine the disguised recognition result.
  • the third recognition sub-module 2043 is used to fuse the salient recognition results and the disguised recognition results according to the salient confidence level and the disguised confidence level to determine the target recognition result.
  • FIG. 10 is a block diagram of another tissue image recognition device according to an exemplary embodiment. As shown in Figure 10, the recognition module 204 may also include:
  • the fourth recognition sub-module 2044 is used to identify the tissue image according to the salient image features to determine the target recognition result if the salience confidence is greater than the first threshold and the camouflage confidence is less than or equal to the second threshold.
  • the fifth identification sub-module 2045 is used to identify the tissue image according to the characteristics of the disguised image to determine the target recognition result if the camouflage confidence is greater than the second threshold and the significant confidence is less than or equal to the first threshold.
  • the sixth identification sub-module 2046 is used to identify the tissue image according to the salient image features to determine the significant recognition result if the significant confidence is less than or equal to the first threshold and the camouflage confidence is less than or equal to the second threshold; according to the camouflage image Features are used to identify tissue images to determine the camouflage recognition results; based on the salience confidence and camouflage confidence, the salient recognition results and camouflage recognition results are fused to determine the target recognition results.
  • the extraction module 202 can be used to:
  • the pre-trained saliency encoder is used to extract features from the tissue image to obtain salient image features
  • the pre-trained camouflage encoder is used to extract features from the tissue image to obtain camouflage image features.
  • the confidence determination module 203 may be used to:
  • the identification module 204 can be used to:
  • the pre-trained decoder is used to identify tissue images based on salient image features, camouflage image features, salience confidence, and camouflage confidence to determine the target recognition results.
  • the salience encoder, camouflage encoder, saliency classifier, camouflage classifier and decoder are jointly trained according to the preset sample image set.
  • the sample image set includes multiple salient sample images and multiple camouflage sample images.
  • the salient sample target in image The object is a salient type
  • the target object in the camouflage sample image is a camouflage type.
  • the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier and decoder are jointly trained in the following way:
  • Step A Obtain a sample input set and a sample output set.
  • the sample input set includes: multiple sample inputs.
  • Each sample input includes a sample image in the sample image set.
  • the sample image is a salient sample image or a camouflage sample image.
  • the sample output set includes sample output corresponding to each sample input.
  • the sample output is the real recognition result of the corresponding sample image. The real recognition result is used to identify the area where the target object is located in the corresponding sample image.
  • Step B input each sample as the input of the salience encoder to obtain the salient sample image features, and input the sample as the input of the camouflage encoder to obtain the camouflage sample image features.
  • Step C Input the salient sample image features into the saliency classifier to obtain the first confidence level, and input the camouflage sample image features into the camouflage classifier to obtain the second confidence level.
  • Step D input the target sample image features into the decoder to obtain the sample recognition result output by the decoder. If the sample input is a significant sample image, the target sample image feature is a significant sample image feature. If the sample input is a camouflage sample image, The target sample image features are the camouflage sample image features.
  • Step E Use the sample recognition result and the sample output corresponding to the sample input as the input of the discriminator to obtain the discrimination result output by the discriminator.
  • Step F Determine the total loss based on the significant sample image features, camouflaged sample image features, first confidence level, second confidence level, sample recognition results, and discrimination results.
  • Step G jointly train the saliency encoder, camouflage encoder, saliency classifier, camouflage classifier, decoder and discriminator according to the total loss.
  • step F can be implemented in the following way:
  • Step F1 Determine the similarity loss based on the salient sample image features and the camouflaged sample image features.
  • Step F2 Determine the confidence loss based on the first confidence level and the second confidence level.
  • Step F3 Determine the recognition loss based on the sample recognition results and the sample output corresponding to each sample input.
  • Step F4 Determine the adversarial loss based on the discrimination results and the sample output corresponding to each sample input.
  • Step F5 Determine the total loss based on similarity loss, confidence loss, recognition loss, and confrontation loss.
  • step F1 may include:
  • the similarity loss is determined based on the cosine similarity of the salient sample image features and the camouflaged sample image features.
  • Step F2 may include:
  • the cross-entropy loss between the first confidence level and the first target true value is summed with the cross-entropy loss between the second confidence level and the second target true value to obtain the confidence loss, the first target true value and the second target true value. Determine based on the corresponding sample input.
  • Step F3 may include:
  • the recognition loss is obtained by summing the cross-entropy loss between the sample recognition result and the sample output corresponding to each sample input, and the Dice loss between the sample recognition result and the sample output corresponding to each sample input.
  • Step F4 may include:
  • the cross-entropy loss between the discriminant result and the third target true value is summed, and the cross-entropy loss between the sample output corresponding to each sample input and the fourth target true value is summed to obtain the adversarial loss, the third target true value and the fourth target true value. Values are determined based on the corresponding sample input.
  • the present disclosure first obtains the tissue images collected by endoscope, and then extracts the significant image features of the tissue images respectively. and camouflage image features, then determine the significant confidence of the tissue image based on the salient image features, determine the camouflage confidence of the tissue image based on the camouflage image features, and finally, combine the salient image features, camouflage image features, significant confidence, and camouflage confidence to Recognize the tissue image to obtain the area where the target object is located in the tissue image.
  • This disclosure respectively extracts the salient image features and camouflaged image features of the tissue image, and further determines the salient confidence and camouflage confidence, thereby determining the area where the target object is located in the tissue image, which can improve the accuracy and generalization ability of image recognition.
  • FIG. 11 it shows a schematic structural diagram of an electronic device (for example, it can be the execution subject in the above embodiment, it can be a terminal device or a server) 300 suitable for implementing embodiments of the present disclosure.
  • Terminal devices in embodiments of the present disclosure may include, but are not limited to, mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablets), PMPs (Portable Multimedia Players), vehicle-mounted terminals (such as Mobile terminals such as car navigation terminals) and fixed terminals such as digital TVs, desktop computers, etc.
  • the electronic device shown in FIG. 11 is only an example and should not impose any limitations on the functions and scope of use of the embodiments of the present disclosure.
  • the electronic device 300 may include a processing device (eg, central processing unit, graphics processor, etc.) 301 , which may be loaded into a random access device according to a program stored in a read-only memory (ROM) 302 or from a storage device 308 .
  • the program in the memory (RAM) 303 executes various appropriate actions and processes.
  • various programs and data required for the operation of the electronic device 300 are also stored.
  • the processing device 301, ROM 302 and RAM 303 are connected to each other via a bus 304.
  • An input/output (I/O) interface 305 is also connected to bus 304.
  • the following devices may be connected to the I/O interface 305: input devices 306 including, for example, a touch screen, touch pad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; including, for example, a liquid crystal display (LCD), speakers, vibration An output device 307 such as a computer; a storage device 308 including a magnetic tape, a hard disk, etc.; and a communication device 309.
  • the communication device 309 may allow the electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data.
  • FIG. 11 illustrates electronic device 300 with various means, it should be understood that implementation or availability of all illustrated means is not required. More or fewer means may alternatively be implemented or provided.
  • embodiments of the present disclosure include a computer program product including a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the method illustrated in the flowchart.
  • the computer program may be downloaded and installed from the network via communication device 309, or from storage device 308, or from ROM 302.
  • the processing device 301 When the computer program is executed by the processing device 301, the above-mentioned functions defined in the method of the embodiment of the present disclosure are performed.
  • the computer-readable medium mentioned above in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two.
  • the computer-readable storage medium may be, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any combination thereof. More specific examples of computer readable storage media may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard drive, random access memory (RAM), read only memory (ROM), removable Programmd read-only memory (EPROM or flash memory), fiber optics, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code therein. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. sequence.
  • Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wire, optical cable, RF (radio frequency), etc., or any suitable combination of the above.
  • terminal devices and servers can communicate using any currently known or future developed network protocol such as HTTP (HyperText Transfer Protocol), and can communicate with digital data in any form or medium.
  • Communications e.g., communications network
  • communications networks include local area networks (“LAN”), wide area networks (“WAN”), the Internet (e.g., the Internet), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or developed in the future network of.
  • the above-mentioned computer-readable medium may be included in the above-mentioned electronic device; it may also exist independently without being assembled into the electronic device.
  • the above-mentioned computer-readable medium carries one or more programs.
  • the electronic device acquires the tissue image collected by the endoscope; respectively extracts the salient images of the tissue image. features and camouflage image features; determine the salient confidence of the tissue image based on the salient image features, and determine the camouflage confidence of the tissue image based on the camouflage image features; based on the salient image features, the camouflage image Features, the salience confidence and the camouflage confidence are used to identify the tissue image to determine a target recognition result, and the target recognition result is used to identify the area where the target object is located in the tissue image.
  • Computer program code for performing the operations of the present disclosure may be written in one or more programming languages, including but not limited to object-oriented programming languages such as Java, Smalltalk, C++, or a combination thereof, Also included are conventional procedural programming languages - such as "C" or similar programming languages.
  • the program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.
  • the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (such as an Internet service provider). connected via the Internet).
  • LAN local area network
  • WAN wide area network
  • Internet service provider such as an Internet service provider
  • each block in the flowchart or block diagram may represent a module, segment, or portion of code that contains one or more logic functions that implement the specified executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown one after another may actually execute substantially in parallel, or they may sometimes execute in the reverse order, depending on the functionality involved.
  • each block of the block diagram and/or flowchart illustration, and combinations of blocks in the block diagram and/or flowchart illustration can be implemented by special purpose hardware-based systems that perform the specified functions or operations. , or can be implemented using a combination of specialized hardware and computer instructions.
  • the modules involved in the embodiments of the present disclosure can be implemented in software or hardware. Among them, the name of the module does not constitute a limitation on the module itself under certain circumstances.
  • the acquisition module can also be described as "a module for acquiring tissue images.”
  • FPGAs Field Programmable Gate Arrays
  • ASICs Application Specific Integrated Circuits
  • ASSPs Application Specific Standard Products
  • SOCs Systems on Chips
  • CPLD Complex Programmable Logical device
  • a machine-readable medium may be a tangible medium that may contain or store a program for use by or in connection with an instruction execution system, apparatus, or device.
  • the machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium.
  • Machine-readable media may include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or devices, or any suitable combination of the foregoing.
  • machine readable storage medium More specific examples of qualities would include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory). flash memory), fiber optics, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
  • Example 1 provides a method for identifying tissue images, including: acquiring tissue images collected by an endoscope; respectively extracting salient image features and disguised image features of the tissue images; according to The salient image features determine the salient confidence of the tissue image, and the camouflage confidence of the tissue image is determined based on the camouflage image features; based on the salient image features, the camouflage image features, and the salient confidence and the camouflage confidence, the tissue image is recognized to determine a target recognition result, and the target recognition result is used to identify the area where the target object is located in the tissue image.
  • Example 2 provides the method of Example 1, wherein based on the salient image features, the camouflage image features, the salience confidence and the camouflage confidence, the Recognizing the tissue image to determine the target recognition result includes: identifying the tissue image according to the salient image features to determine the salient recognition result; identifying the tissue image according to the camouflage image features to determine the camouflage Recognition result; according to the salient confidence level and the camouflage confidence level, fuse the salient recognition result and the camouflage recognition result to determine the target recognition result.
  • Example 3 provides the method of Example 1, wherein based on the salient image features, the camouflage image features, the salience confidence and the camouflage confidence, the Recognizing the tissue image to determine the target recognition result includes: if the significant confidence is greater than a first threshold and the camouflage confidence is less than or equal to a second threshold, identifying the tissue image according to the salient image feature to determine the target recognition result; if the camouflage confidence is greater than the second threshold, and the significant confidence is less than or equal to the first threshold, identify the tissue image according to the camouflage image features to determine the target recognition result; if the significant confidence is less than or equal to the first threshold, and the camouflage confidence is less than or equal to the second threshold, identify the tissue image according to the salient image features to determine the significant recognition result, identify the tissue image according to the disguised image features to determine the disguised recognition result, and compare the significant recognition result and the disguised image according to the significant confidence level and the disguised
  • Example 4 provides the method of Example 1, wherein respectively extracting salient image features and disguised image features of the tissue image includes: using a pre-trained saliency encoder to Perform feature extraction on the image to obtain the salient image features, and use a pre-trained camouflage encoder to perform feature extraction on the tissue image to obtain the camouflage image features; determine the salience of the tissue image based on the salient image features.
  • the method is based on the salient image features, the camouflage image features, the salient confidence and the The camouflage confidence level is used to identify the tissue image to determine the target recognition result, including: using a pre-trained decoder, based on the salient image features, the camouflage image features, the salient confidence level and the camouflage Confidence, the tissue image is identified to determine the target recognition result; the salient encoder, the camouflage encoder, the salient classifier, the camouflage classifier and the decoder, according to the predetermined It is obtained by joint training of a
  • Example 5 provides the method of Example 4, the saliency encoder, the disguise encoder, the saliency classifier, the disguise classifier and the decoder are obtained by Obtained by joint training in the following manner: obtaining a sample input set and a sample output set, the sample input set includes: multiple sample inputs, each of the sample inputs includes the sample A sample image in this image set, the sample image is the salient sample image or the camouflage sample image; the sample output set includes a sample output corresponding to each of the sample inputs, and the sample output is the corresponding The real recognition result of the sample image, the real recognition result is used to identify the area where the target object is located in the corresponding sample image; input each sample as the input of the salient encoder to obtain a salient sample image Features, and input the sample as the input of the camouflage encoder to obtain the camouflage sample image features; input the salient sample image features into the salience classifier to obtain the first confidence, and use the camouflage
  • the target The sample image feature is the significant sample image feature. If the sample input is the camouflage sample image, the target sample image feature is the camouflage sample image feature; the sample recognition result and the sample input corresponding sample output As the input of the discriminator, to obtain the discrimination result output by the discriminator; according to the significant sample image features, the camouflage sample image features, the first confidence level, the second confidence level, the sample identification As a result, the discrimination result determines the total loss; the saliency encoder, the camouflage encoder, the saliency classifier, the camouflage classifier, the decoder and the discriminator are jointly trained according to the total loss .
  • Example 6 provides the method of Example 5, which is based on the salient sample image features, the camouflage sample image features, the first confidence level, the second confidence level Determining the total loss based on the sample identification results and the discrimination results includes: determining the similarity loss based on the significant sample image features and the disguised sample image features; and determining the similarity loss based on the first confidence level and the second confidence level. determine the confidence loss; determine the recognition loss based on the sample recognition result and the sample output corresponding to each sample input; determine the adversarial loss based on the discrimination result and the sample output corresponding to each sample input; The similarity loss, the confidence loss, the recognition loss, and the adversarial loss determine the total loss.
  • Example 7 provides the method of Example 6, wherein determining the similarity loss based on the salient sample image features and the camouflage sample image features includes: based on the salient sample image The cosine similarity between the feature and the camouflage sample image feature determines the similarity loss; determining the confidence loss according to the first confidence level and the second confidence level includes: combining the first confidence level and The cross-entropy loss of the first target true value is summed with the cross-entropy loss of the second confidence and the second target true value to obtain the confidence loss, the first target true value and the second target The true value is determined based on the corresponding sample input; determining the recognition loss based on the sample recognition result and the sample output corresponding to each sample input includes: comparing the sample recognition result with the sample output corresponding to each sample input.
  • the cross-entropy loss of the sample output is summed with the Dice loss of the sample output corresponding to the sample identification result and each of the sample inputs to obtain the identification loss; the discrimination result and each of the sample inputs are Corresponding sample output, determining the adversarial loss includes: summing the cross-entropy loss between the discrimination result and the third target true value, and the cross-entropy loss between the sample output corresponding to each of the sample inputs and the fourth target true value. , the adversarial loss is obtained, and the third target true value and the fourth target true value are determined according to the corresponding sample input.
  • Example 8 provides a device for identifying tissue images, including: an acquisition module for acquiring tissue images collected by an endoscope; and an extraction module for respectively extracting the tissue images.
  • Salient image features and camouflage image features a confidence determination module, configured to determine the salient confidence of the tissue image based on the salient image features, and determine the camouflage confidence of the tissue image based on the camouflage image features;
  • identification A module configured to identify the tissue image according to the salient image features, the camouflage image features, the salient confidence level and the camouflage confidence level to determine a target recognition result, and the target recognition result is used for Identify the region in the tissue image where the target object is located.
  • Example 9 provides a computer-readable medium having a computer program stored thereon, which implements the steps of the methods described in Examples 1 to 7 when executed by a processing device.
  • Example 10 provides an electronic device, including: a storage device storing A computer program is stored; a processing device is configured to execute the computer program in the storage device to implement the steps of the methods described in Examples 1 to 7.
  • Example 11 provides a computer program product, including a computer program that implements the steps of the methods described in Examples 1 to 7 when executed by a processing device.
  • Example 12 provides a computer program that, when executed by a processing device, implements the steps of the methods described in Examples 1 to 7.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

本公开涉及一种组织图像的识别方法、装置、可读介质、电子设备、计算机程序产品以及计算机程序，涉及图像处理技术领域，该方法包括：获取内窥镜采集的组织图像，分别提取组织图像的显著图像特征以及伪装图像特征，根据显著图像特征确定组织图像的显著置信度，并根据伪装图像特征确定组织图像的伪装置信度，根据显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以确定目标识别结果，目标识别结果用于标识组织图像中目标对象所在的区域。本公开分别提取组织图像的显著图像特征和伪装图像特征，进一步确定显著置信度和伪装置信度，以此确定组织图像中目标对象所在的区域，能够提高图像识别的准确度和泛化能力。

Description

组织图像的识别方法、装置、可读介质和电子设备
相关申请的交叉引用
本公开要求于2022年03月30日提交的申请号为202210326583.1、名称为“组织图像的识别方法、装置、可读介质和电子设备”的中国专利申请的优先权，此申请的内容通过引用并入本文。
技术领域
本公开涉及图像处理技术领域，具体地，涉及一种组织图像的识别方法、装置、可读介质、电子设备、计算机程序产品以及计算机程序。
背景技术
内窥镜检查作为医疗领域中一种常见的检查方式，由于其能够直接观察到人体内部的情况，得到了广泛应用。内窥镜检查过程中可以实时采集人体内部的组织图像，进一步的可以通过图像识别的方式，从组织图像中识别出息肉、溃疡等对象所在的区域。以结肠中的息肉为例，图像识别主要针对的是后期息肉，后期息肉的体积较大，比较容易发现，因此识别出息肉的准确率高。然而针对早期息肉，由于早期息肉与结肠背景差别较小、体积较小等原因，识别出息肉的准确度较低，无法满足内窥镜检查的实际需求。
发明内容
提供该发明内容部分以便以简要的形式介绍构思，这些构思将在后面的具体实施方式部分被详细描述。该发明内容部分并不旨在标识要求保护的技术方案的关键特征或必要特征，也不旨在用于限制所要求的保护的技术方案的范围。
第一方面，本公开提供一种组织图像的识别方法，所述方法包括：
获取内窥镜采集的组织图像；
分别提取所述组织图像的显著图像特征以及伪装图像特征；
根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度；
根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，所述目标识别结果用于标识所述组织图像中目标对象所在的区域。
第二方面，本公开提供一种组织图像的识别装置，所述装置包括：
获取模块，用于获取内窥镜采集的组织图像；
提取模块，用于分别提取所述组织图像的显著图像特征以及伪装图像特征；
置信度确定模块，用于根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度；
识别模块，用于根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，所述目标识别结果用于标识所述组织图像中目标对象所在的区域。
第三方面，本公开提供一种计算机可读介质，其上存储有计算机程序，该程序被处理装置执行时实现本公开第一方面所述方法的步骤。
第四方面，本公开提供一种电子设备，包括：
存储装置，其上存储有计算机程序；
处理装置，用于执行所述存储装置中的所述计算机程序，以实现本公开第一方面所述方法的步骤。
第五方面，本公开提供一种计算机程序产品，包括计算机程序，该计算机程序被处理装置执行时实现本公开第一方面所述方法的步骤。
第六方面，本公开提供一种计算机程序，该计算机程序被处理装置执行时实现本公开第一方面所述方法的步骤。
通过上述技术方案，本公开首先获取内窥镜采集的组织图像，然后分别提取组织图像的显著图像特征以及伪装图像特征，再根据显著图像特征确定组织图像的显著置信度，根据伪装图像特征确定组织图像的伪装置信度，最后，结合显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以得到组织图像中目标对象所在的区域。本公开分别提取组织图像的显著图像特征和伪装图像特征，进一步确定显著置信度和伪装置信度，以此确定组织图像中目标对象所在的区域，能够提高图像识别的准确度和泛化能力。
本公开的其他特征和优点将在随后的具体实施方式部分予以详细说明。
附图说明
结合附图并参考以下具体实施方式，本公开各实施例的上述和其他特征、优点及方面将变得更加明显。贯穿附图中，相同或相似的附图标记表示相同或相似的元素。应当理解附图是示意性的，元件和元素不一定按照比例绘制。在附图中：
图1是根据一示例性实施例示出的一种组织图像的识别方法的流程图；
图2是根据一示例性实施例示出的一种识别网络的示意图；
图3是根据一示例性实施例示出的另一种组织图像的识别方法的流程图；
图4是根据一示例性实施例示出的另一种组织图像的识别方法的流程图；
图5是根据一示例性实施例示出的一种显著编码器、伪装编码器、显著分类器、伪装分类器、解码器和判别器的示意图；
图6是根据一示例性实施例示出的一种联合训练显著编码器、伪装编码器、显著分类器、伪装分类器和解码器的流程图；
图7是根据一示例性实施例示出的另一种联合训练显著编码器、伪装编码器、显著分类器、伪装分类器和解码器的流程图；
图8是根据一示例性实施例示出的一种组织图像的识别装置的框图；
图9是根据一示例性实施例示出的另一种组织图像的识别装置的框图；
图10是根据一示例性实施例示出的另一种组织图像的识别装置的框图；
图11是根据一示例性实施例示出的一种电子设备的框图。
具体实施方式
下面将参照附图更详细地描述本公开的实施例。虽然附图中显示了本公开的某些实施例，然 而应当理解的是，本公开可以通过各种形式来实现，而且不应该被解释为限于这里阐述的实施例，相反提供这些实施例是为了更加透彻和完整地理解本公开。应当理解的是，本公开的附图及实施例仅用于示例性作用，并非用于限制本公开的保护范围。
应当理解，本公开的方法实施方式中记载的各个步骤可以按照不同的顺序执行，和/或并行执行。此外，方法实施方式可以包括附加的步骤和/或省略执行示出的步骤。本公开的范围在此方面不受限制。
本文使用的术语“包括”及其变形是开放性包括，即“包括但不限于”。术语“基于”是“至少部分地基于”。术语“一个实施例”表示“至少一个实施例”；术语“另一实施例”表示“至少一个另外的实施例”；术语“一些实施例”表示“至少一些实施例”。其他术语的相关定义将在下文描述中给出。
需要注意，本公开中提及的“第一”、“第二”等概念仅用于对不同的装置、模块或单元进行区分，并非用于限定这些装置、模块或单元所执行的功能的顺序或者相互依存关系。
需要注意，本公开中提及的“一个”、“多个”的修饰是示意性而非限制性的，本领域技术人员应当理解，除非在上下文另有明确指出，否则应该理解为“一个或多个”。
本公开实施方式中的多个装置之间所交互的消息或者信息的名称仅用于说明性的目的，而并不是用于对这些消息或信息的范围进行限制。
本公开中所有获取信号、信息或数据的动作都是在遵照所在地国家相应的数据保护法规政策的前提下，并获得由相应装置所有者给予授权的情况下进行的。
图1是根据一示例性实施例示出的一种组织图像的识别方法的流程图，如图1所示，该方法包括以下步骤：
步骤101，获取内窥镜采集的组织图像。
举例来说，在进行内窥镜检查时，内窥镜会按照预设的采集周期不断地采集人体组织中的图像，即本实施例中的组织图像。组织图像可以为当前时刻内窥镜采集的图像，也可以是任一时刻内窥镜采集的图像。也就是说，组织图像可以是在内窥镜的进镜过程中采集的图像，也可以是在内窥镜的退镜过程中采集的图像，本公开对此不作具体限定。需要说明的是，本公开实施例中所述的内窥镜，例如可以是肠镜、结肠镜、胃镜等，若内窥镜为肠镜，那么上述组织图像即为肠道图像。内窥镜还可以用于采集其他组织的图像，本公开对此不作具体限定。
进一步的，内窥镜检查过程中，可能由于进镜手法不稳定，或者内窥镜所处位置不合适等原因，会采集到很多无效的图像，例如障碍物遮挡、曝光度过大、清晰度过低、体外等图像。这些无效的图像会对图像识别产生干扰，造成处理资源的浪费。因此，在得到组织图像之后，可以先判断组织图像是否有效，以过滤掉无效的组织图像。若组织图像为无效的图像，可以直接丢弃该组织图像，继续获取下一采集周期采集的组织图像。这样能够减少不必要的数据处理，提高处理速度。例如，可以利用预先训练的过滤模型对组织图像进行过滤，以删除无效的组织图像。过滤模型的结构例如可以是CNN(英文：Convolutional Neural Networks，中文：卷积神经网络)或者LSTM(英文：Long Short-Term Memory，中文：长短期记忆网络)，也可以是Transformer(例如Vision Transformer)中的Encoder等，本公开对此不作具体限定。
步骤102，分别提取组织图像的显著图像特征以及伪装图像特征。
示例的，要识别出组织图像中目标对象所在的区域，需要对组织图像进行特征提取，以得到组织图像中包括的图像特征，图像特征可以为特征图(英文：Feature Map)，也可以是特征向量，本公开对此不作具体限定。其中，目标对象可以理解为内窥镜检查过程中需要关注的对象，可以理解为组织上生长出的异物，例如可以是息肉、溃疡等。目标对象的生长过程往往是体积从小到 大，相应的，被识别出的难易程度由大到小，如果能在早期发现目标对象，能够及时对目标对象进行切除或者治疗，避免目标对象进一步恶化。以结肠中的息肉来举例，息肉的发展过程往往是从正常黏膜—黏膜上皮增生—腺瘤性息肉—腺瘤性息肉逐渐增大的过程。
目标对象在早期所具有的特征往往具有伪装特性，不易被识别，即目标对象在早期属于伪装类型的目标对象。而目标对象在后期所具有的特征往往很显著，易于被识别，即目标对象在后期属于显著类型的目标对象。因此可以利用两种不同的特征提取方式，提取组织图像的显著图像特征以及伪装图像特征。其中，提取显著图像特征的方式，可以理解为，适用于组织图像中包括显著类型的目标对象的情况的特征提取方式，提取出的显著图像特征能够反应显著类型的目标对象的特征。提取伪装图像特征的方式，可以理解为，适用于组织图像中包括伪装类型的目标对象的情况的特征提取方式，提取出的伪装图像特征能够反应伪装类型的目标对象的特征。具体的，可以根据包括显著类型的目标对象的样本图像，训练显著提取模型(包括后文提及的显著编码器)，用于提取显著图像特征，并根据包括伪装类型的目标对象的样本图像，训练伪装提取模型(包括后文提及的伪装编码器)，用于提取伪装图像特征。其中，显著提取模型和伪装提取模型的结构可以为：CNN或者VGG(英文：Visual Geometry Group)网络，也可以是Transformer中的Encoder、ResNet、DenseNet等，本公开对此不作具体限定。
步骤103，根据显著图像特征确定组织图像的显著置信度，并根据伪装图像特征确定组织图像的伪装置信度。
示例的，在得到显著图像特征和伪装图像特征之后，可以根据显著图像特征和伪装图像特征来判断组织图像中目标对象是显著类型，还是伪装类型。由于目标对象的显著程度是线性变化，而非二值变化，因此可以通过置信度的方式来表示组织图像中目标对象的显著程度(或者伪装程度)。具体的，可以根据显著图像特征确定组织图像的显著置信度，显著置信度用于表征组织图像中目标对象是显著类型的概率，也可以理解为组织图像中目标对象的显著程度。同时可以根据伪装图像特征确定组织图像的伪装置信度，伪装置信度用于表征组织图像中目标对象是伪装类型的概率，也可以理解为组织图像中目标对象的伪装程度。具体的，可以根据包括显著类型的目标对象的样本图像，训练显著预测模型(包括后文提及的显著分类器)，用于确定显著置信度，并根据包括伪装类型的目标对象的样本图像，训练伪装预测模型，用于确定伪装置信度。其中，显著预测模型和伪装预测模型的结构例如可以是SoftMax模型、决策树模型或者SVM(英文：Support Vector Machine，中文：支持向量机)等，本公开对此不作具体限定。
步骤104，根据显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以确定目标识别结果，目标识别结果用于标识组织图像中目标对象所在的区域。
示例的，可以结合显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以得到目标识别结果，其中，目标识别结果用于指示组织图像中哪些区域属于目标对象。识别过程可以理解为对组织图像进行标注的过程，将组织图像中属于目标对象的像素标注出来，识别过程也可以理解为对组织图像进行分割的过程，将组织图像中属于目标对象的像素提取出来。具体的，目标识别结果可以是与组织图像尺寸相同的二值图像，其中白色区域对应目标对象(或者黑色区域对应目标对象)，目标识别结果也可以是一组坐标范围，用于表示织图像中目标对象的坐标范围。
具体的，可以比较显著置信度和伪装置信度，确定是根据显著图像特征进行识别，还是根据伪装图像特征进行识别。例如，若显著置信度较高，伪装置信度较低，那么说明组织图像中目标对象为显著类型的概率较高，那么可以根据显著图像特征进行识别，以得到目标识别结果。若显 著置信度较低，伪装置信度较高，那么说明组织图像中目标对象为伪装类型的概率较高，那么可以根据伪装图像特征进行识别，以得到目标识别结果。也可以利用显著置信度和伪装置信度对显著图像特征、伪装图像特征进行融合，然后根据融合后的图像特征进行识别，以得到目标识别结果。还可以分别根据显著图像特征进行识别，根据伪装图像特征进行识别，然后利用显著置信度和伪装置信度对两种识别结果进行融合，得到目标识别结果。这样，目标识别结果既考虑了显著图像特征，又考虑了伪装图像特征，能够有效提高对伪装类型的目标对象的识别准确度，从而提高图像识别的泛化能力。
在实现本公开所提供的组织图像的识别方法时，可以利用预先联合训练的显著编码器(表示为Es)、伪装编码器(表示为Ec)、显著分类器(表示为Cs)、伪装分类器(表示为Cc)和解码器(表示为Decoder)来实现。可以将显著编码器、伪装编码器、显著分类器、伪装分类器和解码器理解为一个用于对组织图像进行识别的识别网络，识别网络的结构示意图如图2所示。其中，识别网络根据预设的样本图像集联合训练得到，样本图像集包括多个显著样本图像和多个伪装样本图像，显著样本图像中目标对象为显著类型，伪装样本图像中目标对象为伪装类型。
相应的，步骤102的实现方式可以为：
利用预先训练的显著编码器对组织图像进行特征提取，得到显著图像特征，利用预先训练的伪装编码器对组织图像进行特征提取，得到伪装图像特征。
示例的，可以将组织图像作为显著编码器的输入，显著编码器的输出即为显著图像特征，同时可以将组织图像作为伪装编码器的输入，伪装编码器的数据即为伪装图像特征。其中，显著编码器和伪装编码器均用于实现对图像的特征提取，可以采用ResNet50的结构。
步骤103的实现方式可以为：
将显著图像特征输入预先训练的显著分类器，得到显著分类器输出的显著置信度，将伪装图像特征输入预先训练的伪装分类器，得到伪装分类器输出的伪装置信度。
示例的，可以将显著图像特征输入显著分类器，显著分类器输出的即为显著置信度(表示为salient confident)，同时可以将伪装图像特征输入伪装分类器，伪装分类器输出的即为伪装置信度(表示为camouflaged confident)。其中显著分类器用于确定组织图像与显著类型的匹配概率，即显著置信度。伪装分类器用于确定组织图像与伪装类型的匹配概率，即伪装置信度。
步骤104的实现方式可以为：
利用预先训练的解码器，根据显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以确定目标识别结果。
示例的，可以利用解码器，结合显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，得到目标识别结果。以下对利用解码器对组织图像进行识别的实现方式进行具体说明。
图3是根据一示例性实施例示出的另一种组织图像的识别方法的流程图，如图3所示，步骤104的实现方式可以包括：
步骤1041，根据显著图像特征对组织图像进行识别，以确定显著识别结果。
步骤1042，根据伪装图像特征对组织图像进行识别，以确定伪装识别结果。
步骤1043，根据显著置信度和伪装置信度，对显著识别结果和伪装识别结果进行融合，以确定目标识别结果。
示例的，可以将显著图像特征输入解码器，以实现对组织图像的识别，此时解码器输出显著识别结果。可以将伪装图像特征输入解码器，以实现对组织图像的识别，此时解码器输出伪装识 别结果。
之后，可以根据显著置信度和伪装置信度，对解码器输出的显著识别结果和伪装识别结果进行融合，并将融合结果作为目标识别结果。具体的，可以将显著置信度作为显著识别结果的权重，将伪装置信度作为伪装识别结果的权重，以对显著识别结果和伪装识别结果进行加权求和，得到的结果即为目标识别结果。
图4是根据一示例性实施例示出的另一种组织图像的识别方法的流程图，如图4所示，步骤104的实现方式也可以包括：
步骤1044，若显著置信度大于第一阈值，且伪装置信度小于或等于第二阈值，根据显著图像特征对组织图像进行识别，以确定目标识别结果。
步骤1045，若伪装置信度大于第二阈值，且显著置信度小于或等于第一阈值，根据伪装图像特征对组织图像进行识别，以确定目标识别结果。
步骤1046，若显著置信度小于或等于第一阈值，且伪装置信度小于或等于第二阈值，根据显著图像特征对组织图像进行识别，以确定显著识别结果；根据伪装图像特征对组织图像进行识别，以确定伪装识别结果；根据显著置信度和伪装置信度，对显著识别结果和伪装识别结果进行融合，以确定目标识别结果。
示例的，可以先将显著置信度与预设的第一阈值比较，同时将伪装置信度与预设的第二阈值比较，然后根据比较结果确定如何对组织图像进行识别，第一阈值和第二阈值可以相同或不同。例如第一阈值可以为0.7，第二阈值可以为0.6。
在显著置信度大于第一阈值，且伪装置信度小于或等于第二阈值的情况下，说明组织图像中目标对象为显著类型的概率较高，那么可以将显著图像特征输入解码器，以实现对组织图像的识别，此时解码器输出的即为目标识别结果。在伪装置信度大于第二阈值，且显著置信度小于或等于第一阈值的情况下，说明组织图像中目标对象为伪装类型的概率较高，那么可以将伪装图像特征输入解码器，以实现对组织图像的识别，此时解码器输出的即为目标识别结果。
在显著置信度小于或等于第一阈值，且伪装置信度小于或等于第二阈值的情况下，说明组织图像中目标对象为显著类型或者伪装类型的概率都不高，那么可以将显著图像特征输入解码器，此时解码器输出显著识别结果。再将伪装图像特征输入解码器，此时解码器输出伪装识别结果。之后，可以根据显著置信度和伪装置信度，对解码器输出的显著识别结果和伪装识别结果进行融合，并将融合结果作为目标识别结果。具体的，可以将显著置信度作为显著识别结果的权重，将伪装置信度作为伪装识别结果的权重，以对显著识别结果和伪装识别结果进行加权求和，得到的结果即为目标识别结果。
在对显著编码器、伪装编码器、显著分类器、伪装分类器和解码器进行联合训练时，可以与预设的判别器一起进行训练，显著编码器、伪装编码器、显著分类器、伪装分类器、解码器和判别器的连接关系如图5所示。
图6是根据一示例性实施例示出的一种联合训练显著编码器、伪装编码器、显著分类器、伪装分类器和解码器的流程图，如图6所示，显著编码器、伪装编码器、显著分类器、伪装分类器和解码器是通过以下方式联合训练得到的：
步骤A，获取样本输入集和样本输出集，样本输入集包括：多个样本输入，每个样本输入包括样本图像集中的一个样本图像，样本图像为显著样本图像或伪装样本图像。样本输出集中包括与每个样本输入对应的样本输出，样本输出为对应的样本图像的真实识别结果，真实识别结果用于标识对应的样本图像中目标对象所在的区域。
举例来说，在对显著编码器、伪装编码器、显著分类器、伪装分类器和解码器进行联合训练时，需要先获取用于训练的样本输入集和样本输出集。其中，样本输入集包括：多个样本输入，每个样本输入包括样本图像集中的一个样本图像，样本图像可以为显著样本图像(表示为salient data)，也可以为伪装样本图像(表示为camouflaged data)。样本输出集中包括与每个样本输入对应的样本输出，样本输出为对应的样本图像的真实识别结果，真实识别结果用于表示对应的样本图像中目标对象所在的区域，真实识别结果可以是二值图像，也可以是坐标范围。样本图像可以是之前执行内窥镜检查时采集到的大量的组织图像。可以将样本输入集和样本输出集中，按照60％、10％、30％的比例分为训练集、验证集和测试集，本公开对比例分配不作具体限定。
步骤B，将每个样本输入作为显著编码器的输入，得到显著样本图像特征，并将该样本输入作为伪装编码器的输入，得到伪装样本图像特征。
步骤C，将显著样本图像特征输入显著分类器，得到第一置信度，并将伪装样本图像特征输入伪装分类器，得到第二置信度。
示例的，针对每个样本输入，将该样本输入分别作为显著编码器和伪装编码器的输入，得到显著编码器输出的显著样本图像特征，以及伪装编码器输出的伪装样本图像特征。之后，可以将显著样本图像特征输入显著分类器，得到显著分类器输出的第一置信度，并将伪装样本图像特征输入伪装分类器，得到伪装分类器输出的第二置信度。其中，第一置信度用于表征该样本图像中目标对象是显著类型的概率，也可以理解为该样本图像中目标对象的显著程度。第二置信度用于表征该样本图像中目标对象是伪装类型的概率，也可以理解为该样本图像中目标对象的伪装程度。
步骤D，将目标样本图像特征输入解码器，以得到解码器输出的样本识别结果，若该样本输入为显著样本图像，目标样本图像特征为显著样本图像特征，若该样本输入为伪装样本图像，目标样本图像特征为伪装样本图像特征。
步骤E，将样本识别结果和该样本输入对应的样本输出作为判别器的输入，以得到判别器输出的判别结果。
示例的，根据该样本输入为显著样本图像或者伪装样本图像，确定目标样本图像特征，然后将目标样本图像特征输入解码器，以得到解码器输出的样本识别结果，样本识别结果用于表示解码器预测该样本图像中目标对象所在的区域，可以是二值图像，也可以是坐标范围。具体的，若该样本输入为显著样本图像，目标样本图像特征为显著样本图像特征，若该样本输入为伪装样本图像，目标样本图像特征为伪装样本图像特征。
之后，可以将样本识别结果和该样本输入对应的样本输出作为判别器(表示为Discriminator)的输入，以得到判别器输出的判别结果。判别结果用于指示输入为显著类型的不确定性(表示为salient uncertainty)，或者为伪装类型的不确定性(表示为camouflaged uncertainty)。具体的，将样本识别结果输入判别器，判别器输出第一判别结果，将该样本输入对应的样本输出输入判别器，判别器输出第二判别结果，训练的过程中让第一判别结果尽可能与第二判别结果不同，以达到对抗训练的目的。在该样本输入为显著样本图像的情况下，第一判别结果用于指示样本识别结果与显著类型的匹配概率，第二判别结果用于指示对应的样本输出与显著类型的匹配概率。在该样本输入为伪装样本图像的情况下，第一判别结果用于指示样本识别结果与伪装类型的匹配概率，第二判别结果用于指示对应的样本输出与伪装类型的匹配概率。其中，判别器可以是GAN(英文：Generative Adversarial Networks，中文：生成式对抗网络)中的判别器，结构可以是ResNet34。
步骤F，根据显著样本图像特征、伪装样本图像特征、第一置信度、第二置信度、样本识别结果、判别结果确定总损失。
步骤G，根据总损失联合训练显著编码器、伪装编码器、显著分类器、伪装分类器、解码器和判别器。
示例的，可以根据显著样本图像特征、伪装样本图像特征、第一置信度、第二置信度、样本识别结果、判别结果确定总损失。之后以降低总损失为目标，利用反向传播算法联合训练显著编码器、伪装编码器、显著分类器、伪装分类器、解码器和判别器。
进一步的，联合训练显著编码器、伪装编码器、显著分类器、伪装分类器、解码器和判别器的初始学习率可以设置为：1e-4，Batch size可以设置为：32，优化器可以选择：Adam，Epoch可以设置为：100，样本图像的大小可以为：256×256。
图7是根据一示例性实施例示出的另一种联合训练显著编码器、伪装编码器、显著分类器、伪装分类器和解码器的流程图，如图7所示，步骤F可以通过以下方式来实现：
步骤F1，根据显著样本图像特征和伪装样本图像特征，确定相似度损失。
举例来说，可以根据显著样本图像特征和伪装样本图像特征的余弦相似度确定相似度损失。例如可以通过公式1来确定相似度损失：
Ls＝Ss+Sc

其中，Ls表示相似度损失，Ss表示样本输入为显著样本图像时，显著样本图像特征和伪装样本图像特征的余弦相似度，表示样本输入为显著样本图像时Es的输出，即显著样本图像对应的显著样本图像特征，表示样本输入为显著样本图像时Ec的输出，即显著样本图像对应的伪装样本图像特征。Sc表示样本输入为伪装样本图像时，显著样本图像特征和伪装样本图像特征的余弦相似度，表示样本输入为伪装样本图像时Es的输出，即伪装样本图像对应的显著样本图像特征，表示样本输入为伪装样本图像时Ec的输出，即伪装样本图像对应的伪装样本图像特征，|| ||表示范数。
步骤F2，根据第一置信度和第二置信度确定置信度损失。
示例的，可以将第一置信度和第一目标真值的交叉熵损失，与第二置信度和第二目标真值的交叉熵损失求和，得到置信度损失，第一目标真值和第二目标真值根据对应的样本输入确定。例如可以通过公式2来确定置信度损失：
Lc＝fce[(Cs(xs)，1)，(Cs(xc)，0)]+fce[(Cc(xs)，0)，(Cc(xc)，1)]      公式2
其中，Lc表示置信度损失，fce表示交叉熵(英文：Cross Entropy)损失，Cs()表示显著分类器的处理，Cc()表示伪装分类器的处理，xs表示显著样本图像，xc表示伪装样本图像。
Cs(xs)表示样本输入为显著样本图像时，Cs的输出，即显著样本图像对应的第一置信度，此时第一目标真值即为1，也就是说当样本输入为显著样本图像时，期望Cs的输出为1。Cs(xc)表示样本输入为伪装样本图像时，Cs的输出，即伪装样本图像对应的第一置信度，此时第一目标真值即为0，也就是说当样本输入为伪装样本图像时，期望Cs的输出为0。
同样的，Cc(xs)表示样本输入为显著样本图像时，Cc的输出，即显著样本图像对应的第二置信度，此时第二目标真值即为0，也就是说当样本输入为显著样本图像时，期望Cc的输出为0。Cc(xc)表示样本输入为伪装样本图像时，Cc的输出，即伪装样本图像对应的第二置信度，此时第二目标真值即为1，也就是说当样本输入为伪装样本图像时，期望Cc的输出为1。
步骤F3，根据样本识别结果和每个样本输入对应的样本输出，确定识别损失。
示例的，可以将样本识别结果与每个样本输入对应的样本输出的交叉熵损失，与样本识别结果与每个样本输入对应的样本输出的Dice损失求和，得到识别损失。例如可以通过公式3来确定识别损失：
其中，Lp表示识别损失，fbce表示二分类交叉熵(英文：Binary Cross Entropy)损失，fdice表示Dice损失，表示样本输入为显著样本图像时Decoder的输出，即显著样本图像对应的样本识别结果，Ys表示显著样本图像对应的样本输出。表示样本输入为伪装样本图像时Decoder的输出，即伪装样本图像对应的样本识别结果，Yc表示伪装样本图像对应的样本输出。
步骤F4，根据判别结果和每个样本输入对应的样本输出，确定对抗损失。
示例的，可以将判别结果和第三目标真值的交叉熵损失，与每个样本输入对应的样本输出和第四目标真值的交叉熵损失求和，得到对抗损失，第三目标真值和第四目标真值根据对应的样本输入确定。例如可以通过公式4来确定对抗损失：
其中，LAD表示对抗损失，fbce表示二分类交叉熵损失，D()表示判别器的处理。表示样本输入为显著样本图像时Decoder的输出，即显著样本图像对应的样本识别结果，Ys表示显著样本图像对应的样本输出，表示显著样本图像对应的样本识别结果对应的判别结果，此时，第三目标真值为0，也就是说将样本识别结果输入Discriminator时，期望Discriminator的输出为0(判别为假)。D(Ys)表示显著样本图像对应的样本输出对应的判别结果，此时第四目标真值为1，也就是说，将显著样本图像对应的样本输出输入Discriminator时，期望Discriminator的输出为1(判别为真)。
表示样本输入为伪装样本图像时Decoder的输出，即伪装样本图像对应的样本识别结果，Yc表示伪装样本图像对应的样本输出。表示伪装样本图像对应的样本识别结果对应的判别结果，此时，第三目标真值为0，也就是说将样本识别结果输入Discriminator时，期望Discriminator的输出为0(判别为假)。D(Yc)表示伪装样本图像对应的样本输出对应的判别结果，此时第四目标真值为1，也就是说，将伪装样本图像对应的样本输出输入Discriminator时，期望Discriminator的输出为1(判别为真)。
步骤F5，根据相似度损失、置信度损失、识别损失、对抗损失，确定总损失。
示例的，可以对相似度损失、置信度损失、识别损失、对抗损失求和，得到总损失，或者对相似度损失、置信度损失、识别损失、对抗损失进行加权求和，得到总损失。例如可以通过公式5来确定总损失：
L＝Ls+Lc+Lp+LAD    公式5
其中，L表示总损失。
综上所述，本公开首先获取内窥镜采集的组织图像，然后分别提取组织图像的显著图像特征 以及伪装图像特征，再根据显著图像特征确定组织图像的显著置信度，根据伪装图像特征确定组织图像的伪装置信度，最后，结合显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以得到组织图像中目标对象所在的区域。本公开分别提取组织图像的显著图像特征和伪装图像特征，进一步确定显著置信度和伪装置信度，以此确定组织图像中目标对象所在的区域，能够提高图像识别的准确度和泛化能力。
图8是根据一示例性实施例示出的一种组织图像的识别装置的框图，如图8所示，该装置200可以包括：
获取模块201，用于获取内窥镜采集的组织图像。
提取模块202，用于分别提取组织图像的显著图像特征以及伪装图像特征。
置信度确定模块203，用于根据显著图像特征确定组织图像的显著置信度，并根据伪装图像特征确定组织图像的伪装置信度。
识别模块204，用于根据显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以确定目标识别结果，目标识别结果用于标识组织图像中目标对象所在的区域。
图9是根据一示例性实施例示出的另一种组织图像的识别装置的框图，如图9所示，识别模块204可以包括：
第一识别子模块2041，用于根据显著图像特征对组织图像进行识别，以确定显著识别结果。
第二识别子模块2042，用于根据伪装图像特征对组织图像进行识别，以确定伪装识别结果。
第三识别子模块2043，用于根据显著置信度和伪装置信度，对显著识别结果和伪装识别结果进行融合，以确定目标识别结果。
图10是根据一示例性实施例示出的另一种组织图像的识别装置的框图，如图10所示，识别模块204还可以包括：
第四识别子模块2044，用于若显著置信度大于第一阈值，且伪装置信度小于或等于第二阈值，根据显著图像特征对组织图像进行识别，以确定目标识别结果。
第五识别子模块2045，用于若伪装置信度大于第二阈值，且显著置信度小于或等于第一阈值，根据伪装图像特征对组织图像进行识别，以确定目标识别结果。
第六识别子模块2046，用于若显著置信度小于或等于第一阈值，且伪装置信度小于或等于第二阈值，根据显著图像特征对组织图像进行识别，以确定显著识别结果；根据伪装图像特征对组织图像进行识别，以确定伪装识别结果；根据显著置信度和伪装置信度，对显著识别结果和伪装识别结果进行融合，以确定目标识别结果。
在一种实现方式中，提取模块202可以用于：
利用预先训练的显著编码器对组织图像进行特征提取，得到显著图像特征，利用预先训练的伪装编码器对组织图像进行特征提取，得到伪装图像特征。
置信度确定模块203可以用于：
将显著图像特征输入预先训练的显著分类器，得到显著分类器输出的显著置信度，将伪装图像特征输入预先训练的伪装分类器，得到伪装分类器输出的伪装置信度。
识别模块204可以用于：
利用预先训练的解码器，根据显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以确定目标识别结果。
其中，显著编码器、伪装编码器、显著分类器、伪装分类器和解码器，根据预设的样本图像集联合训练得到，样本图像集包括多个显著样本图像和多个伪装样本图像，显著样本图像中目标 对象为显著类型，伪装样本图像中目标对象为伪装类型。
在一种实现方式中，显著编码器、伪装编码器、显著分类器、伪装分类器和解码器是通过以下方式联合训练得到的：
步骤A，获取样本输入集和样本输出集，样本输入集包括：多个样本输入，每个样本输入包括样本图像集中的一个样本图像，样本图像为显著样本图像或伪装样本图像。样本输出集中包括与每个样本输入对应的样本输出，样本输出为对应的样本图像的真实识别结果，真实识别结果用于标识对应的样本图像中目标对象所在的区域。
步骤B，将每个样本输入作为显著编码器的输入，得到显著样本图像特征，并将该样本输入作为伪装编码器的输入，得到伪装样本图像特征。
步骤C，将显著样本图像特征输入显著分类器，得到第一置信度，并将伪装样本图像特征输入伪装分类器，得到第二置信度。
步骤D，将目标样本图像特征输入解码器，以得到解码器输出的样本识别结果，若该样本输入为显著样本图像，目标样本图像特征为显著样本图像特征，若该样本输入为伪装样本图像，目标样本图像特征为伪装样本图像特征。
步骤E，将样本识别结果和该样本输入对应的样本输出作为判别器的输入，以得到判别器输出的判别结果。
步骤F，根据显著样本图像特征、伪装样本图像特征、第一置信度、第二置信度、样本识别结果、判别结果确定总损失。
步骤G，根据总损失联合训练显著编码器、伪装编码器、显著分类器、伪装分类器、解码器和判别器。
在另一种实现方式中，步骤F可以通过以下方式来实现：
步骤F1，根据显著样本图像特征和伪装样本图像特征，确定相似度损失。
步骤F2，根据第一置信度和第二置信度确定置信度损失。
步骤F3，根据样本识别结果和每个样本输入对应的样本输出，确定识别损失。
步骤F4，根据判别结果和每个样本输入对应的样本输出，确定对抗损失。
步骤F5，根据相似度损失、置信度损失、识别损失、对抗损失，确定总损失。
在又一种实现方式中，步骤F1可以包括：
根据显著样本图像特征和伪装样本图像特征的余弦相似度确定相似度损失。
步骤F2可以包括：
将第一置信度和第一目标真值的交叉熵损失，与第二置信度和第二目标真值的交叉熵损失求和，得到置信度损失，第一目标真值和第二目标真值根据对应的样本输入确定。
步骤F3可以包括：
将样本识别结果与每个样本输入对应的样本输出的交叉熵损失，与样本识别结果与每个样本输入对应的样本输出的Dice损失求和，得到识别损失。
步骤F4可以包括：
将判别结果和第三目标真值的交叉熵损失，与每个样本输入对应的样本输出和第四目标真值的交叉熵损失求和，得到对抗损失，第三目标真值和第四目标真值根据对应的样本输入确定。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
综上所述，本公开首先获取内窥镜采集的组织图像，然后分别提取组织图像的显著图像特征 以及伪装图像特征，再根据显著图像特征确定组织图像的显著置信度，根据伪装图像特征确定组织图像的伪装置信度，最后，结合显著图像特征、伪装图像特征、显著置信度和伪装置信度，对组织图像进行识别，以得到组织图像中目标对象所在的区域。本公开分别提取组织图像的显著图像特征和伪装图像特征，进一步确定显著置信度和伪装置信度，以此确定组织图像中目标对象所在的区域，能够提高图像识别的准确度和泛化能力。
下面参考图11，其示出了适于用来实现本公开实施例的电子设备(例如可以上述实施例中的执行主体，可以是终端设备或服务器)300的结构示意图。本公开实施例中的终端设备可以包括但不限于诸如移动电话、笔记本电脑、数字广播接收器、PDA(个人数字助理)、PAD(平板电脑)、PMP(便携式多媒体播放器)、车载终端(例如车载导航终端)等等的移动终端以及诸如数字TV、台式计算机等等的固定终端。图11示出的电子设备仅仅是一个示例，不应对本公开实施例的功能和使用范围带来任何限制。
如图11所示，电子设备300可以包括处理装置(例如中央处理器、图形处理器等)301，其可以根据存储在只读存储器(ROM)302中的程序或者从存储装置308加载到随机访问存储器(RAM)303中的程序而执行各种适当的动作和处理。在RAM 303中，还存储有电子设备300操作所需的各种程序和数据。处理装置301、ROM 302以及RAM 303通过总线304彼此相连。输入/输出(I/O)接口305也连接至总线304。
通常，以下装置可以连接至I/O接口305：包括例如触摸屏、触摸板、键盘、鼠标、摄像头、麦克风、加速度计、陀螺仪等的输入装置306；包括例如液晶显示器(LCD)、扬声器、振动器等的输出装置307；包括例如磁带、硬盘等的存储装置308；以及通信装置309。通信装置309可以允许电子设备300与其他设备进行无线或有线通信以交换数据。虽然图11示出了具有各种装置的电子设备300，但是应理解的是，并不要求实施或具备所有示出的装置。可以替代地实施或具备更多或更少的装置。
特别地，根据本公开的实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括承载在非暂态计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信装置309从网络上被下载和安装，或者从存储装置308被安装，或者从ROM302被安装。在该计算机程序被处理装置301执行时，执行本公开实施例的方法中限定的上述功能。
需要说明的是，本公开上述的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本公开中，计算机可读信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读信号介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程 序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：电线、光缆、RF(射频)等等，或者上述的任意合适的组合。
在一些实施方式中，终端设备、服务器可以利用诸如HTTP(HyperText Transfer Protocol，超文本传输协议)之类的任何当前已知或未来研发的网络协议进行通信，并且可以与任意形式或介质的数字数据通信(例如，通信网络)互连。通信网络的示例包括局域网(“LAN”)，广域网(“WAN”)，网际网(例如，互联网)以及端对端网络(例如，ad hoc端对端网络)，以及任何当前已知或未来研发的网络。
上述计算机可读介质可以是上述电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。
上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被该电子设备执行时，使得该电子设备：获取内窥镜采集的组织图像；分别提取所述组织图像的显著图像特征以及伪装图像特征；根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度；根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，所述目标识别结果用于标识所述组织图像中目标对象所在的区域。
可以以一种或多种程序设计语言或其组合来编写用于执行本公开的操作的计算机程序代码，上述程序设计语言包括但不限于面向对象的程序设计语言——诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言——诸如“C”语言或类似的程序设计语言。程序代码可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络——包括局域网(LAN)或广域网(WAN)——连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。
附图中的流程图和框图，图示了按照本公开各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，该模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
描述于本公开实施例中所涉及到的模块可以通过软件的方式实现，也可以通过硬件的方式来实现。其中，模块的名称在某种情况下并不构成对该模块本身的限定，例如，获取模块还可以被描述为“获取组织图像的模块”。
本文中以上描述的功能可以至少部分地由一个或多个硬件逻辑部件来执行。例如，非限制性地，可以使用的示范类型的硬件逻辑部件包括：现场可编程门阵列(FPGA)、专用集成电路(ASIC)、专用标准产品(ASSP)、片上系统(SOC)、复杂可编程逻辑设备(CPLD)等等。
在本公开的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介 质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦除可编程只读存储器(EPROM或快闪存储器)、光纤、便捷式紧凑盘只读存储器(CD-ROM)、光学储存设备、磁储存设备、或上述内容的任何合适组合。
根据本公开的一个或多个实施例，示例1提供了一种组织图像的识别方法，包括：获取内窥镜采集的组织图像；分别提取所述组织图像的显著图像特征以及伪装图像特征；根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度；根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，所述目标识别结果用于标识所述组织图像中目标对象所在的区域。
根据本公开的一个或多个实施例，示例2提供了示例1的方法，所述根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，包括：根据所述显著图像特征对所述组织图像进行识别，以确定显著识别结果；根据所述伪装图像特征对所述组织图像进行识别，以确定伪装识别结果；根据所述显著置信度和所述伪装置信度，对所述显著识别结果和所述伪装识别结果进行融合，以确定所述目标识别结果。
根据本公开的一个或多个实施例，示例3提供了示例1的方法，所述根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，包括：若所述显著置信度大于第一阈值，且所述伪装置信度小于或等于第二阈值，根据所述显著图像特征对所述组织图像进行识别，以确定所述目标识别结果；若所述伪装置信度大于所述第二阈值，且所述显著置信度小于或等于所述第一阈值，根据所述伪装图像特征对所述组织图像进行识别，以确定所述目标识别结果；若所述显著置信度小于或等于所述第一阈值，且所述伪装置信度小于或等于第二阈值，根据所述显著图像特征对所述组织图像进行识别，以确定显著识别结果，根据所述伪装图像特征对所述组织图像进行识别，以确定伪装识别结果，根据所述显著置信度和所述伪装置信度，对所述显著识别结果和所述伪装识别结果进行融合，以确定所述目标识别结果。
根据本公开的一个或多个实施例，示例4提供了示例1的方法，所述分别提取所述组织图像的显著图像特征以及伪装图像特征，包括：利用预先训练的显著编码器对所述组织图像进行特征提取，得到所述显著图像特征，利用预先训练的伪装编码器对所述组织图像进行特征提取，得到所述伪装图像特征；所述根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度，包括：将所述显著图像特征输入预先训练的显著分类器，得到所述显著分类器输出的所述显著置信度，将所述伪装图像特征输入预先训练的伪装分类器，得到所述伪装分类器输出的所述伪装置信度；所述根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，包括：利用预先训练的解码器，根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定所述目标识别结果；所述显著编码器、所述伪装编码器、所述显著分类器、所述伪装分类器和所述解码器，根据预设的样本图像集联合训练得到，所述样本图像集包括多个显著样本图像和多个伪装样本图像，所述显著样本图像中所述目标对象为显著类型，所述伪装样本图像中所述目标对象为伪装类型。
根据本公开的一个或多个实施例，示例5提供了示例4的方法，所述显著编码器、所述伪装编码器、所述显著分类器、所述伪装分类器和所述解码器是通过以下方式联合训练得到的：获取样本输入集和样本输出集，所述样本输入集包括：多个样本输入，每个所述样本输入包括所述样 本图像集中的一个样本图像，所述样本图像为所述显著样本图像或所述伪装样本图像；所述样本输出集中包括与每个所述样本输入对应的样本输出，所述样本输出为对应的所述样本图像的真实识别结果，所述真实识别结果用于标识对应的所述样本图像中目标对象所在的区域；将每个所述样本输入作为所述显著编码器的输入，得到显著样本图像特征，并将该样本输入作为所述伪装编码器的输入，得到伪装样本图像特征；将所述显著样本图像特征输入所述显著分类器，得到第一置信度，并将所述伪装样本图像特征输入所述伪装分类器，得到第二置信度；将目标样本图像特征输入所述解码器，以得到所述解码器输出的样本识别结果，若该样本输入为所述显著样本图像，所述目标样本图像特征为所述显著样本图像特征，若该样本输入为所述伪装样本图像，所述目标样本图像特征为所述伪装样本图像特征；将所述样本识别结果和该样本输入对应的样本输出作为判别器的输入，以得到所述判别器输出的判别结果；根据所述显著样本图像特征、所述伪装样本图像特征、所述第一置信度、所述第二置信度、所述样本识别结果、所述判别结果确定总损失；根据所述总损失联合训练所述显著编码器、所述伪装编码器、所述显著分类器、所述伪装分类器、所述解码器和所述判别器。
根据本公开的一个或多个实施例，示例6提供了示例5的方法，所述根据所述显著样本图像特征、所述伪装样本图像特征、所述第一置信度、所述第二置信度、所述样本识别结果、所述判别结果确定总损失，包括：根据所述显著样本图像特征和所述伪装样本图像特征，确定相似度损失；根据所述第一置信度和所述第二置信度确定置信度损失；根据所述样本识别结果和每个所述样本输入对应的样本输出，确定识别损失；根据所述判别结果和每个所述样本输入对应的样本输出，确定对抗损失；根据所述相似度损失、所述置信度损失、所述识别损失、所述对抗损失，确定所述总损失。
根据本公开的一个或多个实施例，示例7提供了示例6的方法，所述根据所述显著样本图像特征和所述伪装样本图像特征，确定相似度损失，包括：根据所述显著样本图像特征和所述伪装样本图像特征的余弦相似度确定所述相似度损失；所述根据所述第一置信度和所述第二置信度确定置信度损失，包括：将所述第一置信度和第一目标真值的交叉熵损失，与所述第二置信度和第二目标真值的交叉熵损失求和，得到所述置信度损失，所述第一目标真值和所述第二目标真值根据对应的样本输入确定；所述根据所述样本识别结果和每个所述样本输入对应的样本输出，确定识别损失，包括：将所述样本识别结果与每个所述样本输入对应的样本输出的交叉熵损失，与所述样本识别结果与每个所述样本输入对应的样本输出的Dice损失求和，得到所述识别损失；所述根据所述判别结果和每个所述样本输入对应的样本输出，确定对抗损失，包括：将所述判别结果和第三目标真值的交叉熵损失，与每个所述样本输入对应的样本输出和第四目标真值的交叉熵损失求和，得到所述对抗损失，所述第三目标真值和所述第四目标真值根据对应的样本输入确定。
根据本公开的一个或多个实施例，示例8提供了一种组织图像的识别装置，包括：获取模块，用于获取内窥镜采集的组织图像；提取模块，用于分别提取所述组织图像的显著图像特征以及伪装图像特征；置信度确定模块，用于根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度；识别模块，用于根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，所述目标识别结果用于标识所述组织图像中目标对象所在的区域。
根据本公开的一个或多个实施例，示例9提供了一种计算机可读介质，其上存储有计算机程序，该程序被处理装置执行时实现示例1至示例7中所述方法的步骤。
根据本公开的一个或多个实施例，示例10提供了一种电子设备，包括：存储装置，其上存 储有计算机程序；处理装置，用于执行所述存储装置中的所述计算机程序，以实现示例1至示例7中所述方法的步骤。
根据本公开的一个或多个实施例，示例11提供了一种计算机程序产品，包括计算机程序，所述计算机程序被处理装置执行时实现示例1至示例7中所述方法的步骤。
根据本公开的一个或多个实施例，示例12提供了一种计算机程序，所述计算机程序被处理装置执行时实现示例1至示例7中所述方法的步骤。
以上描述仅为本公开的较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本公开中所涉及的公开范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离上述公开构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本公开中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。
此外，虽然采用特定次序描绘了各操作，但是这不应当理解为要求这些操作以所示出的特定次序或以顺序次序执行来执行。在一定环境下，多任务和并行处理可能是有利的。同样地，虽然在上面论述中包含了若干具体实现细节，但是这些不应当被解释为对本公开的范围的限制。在单独的实施例的上下文中描述的某些特征还可以组合地实现在单个实施例中。相反地，在单个实施例的上下文中描述的各种特征也可以单独地或以任何合适的子组合的方式实现在多个实施例中。
尽管已经采用特定于结构特征和/或方法逻辑动作的语言描述了本主题，但是应当理解所附权利要求书中所限定的主题未必局限于上面描述的特定特征或动作。相反，上面所描述的特定特征和动作仅仅是实现权利要求书的示例形式。关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。

Claims (12)

  1. 一种组织图像的识别方法，其中，所述方法包括：
    获取内窥镜采集的组织图像；
    分别提取所述组织图像的显著图像特征以及伪装图像特征；
    根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度；
    根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，所述目标识别结果用于标识所述组织图像中目标对象所在的区域。
  2. 根据权利要求1所述的方法，其中，所述根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，包括：
    根据所述显著图像特征对所述组织图像进行识别，以确定显著识别结果；
    根据所述伪装图像特征对所述组织图像进行识别，以确定伪装识别结果；
    根据所述显著置信度和所述伪装置信度，对所述显著识别结果和所述伪装识别结果进行融合，以确定所述目标识别结果。
  3. 根据权利要求1所述的方法，其中，所述根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，包括：
    若所述显著置信度大于第一阈值，且所述伪装置信度小于或等于第二阈值，根据所述显著图像特征对所述组织图像进行识别，以确定所述目标识别结果；
    若所述伪装置信度大于所述第二阈值，且所述显著置信度小于或等于所述第一阈值，根据所述伪装图像特征对所述组织图像进行识别，以确定所述目标识别结果；
    若所述显著置信度小于或等于所述第一阈值，且所述伪装置信度小于或等于第二阈值，根据所述显著图像特征对所述组织图像进行识别，以确定显著识别结果；根据所述伪装图像特征对所述组织图像进行识别，以确定伪装识别结果；根据所述显著置信度和所述伪装置信度，对所述显著识别结果和所述伪装识别结果进行融合，以确定所述目标识别结果。
  4. 根据权利要求1至3中任一项所述的方法，其中，所述分别提取所述组织图像的显著图像特征以及伪装图像特征，包括：
    利用预先训练的显著编码器对所述组织图像进行特征提取，得到所述显著图像特征，利用预先训练的伪装编码器对所述组织图像进行特征提取，得到所述伪装图像特征；
    所述根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度，包括：
    将所述显著图像特征输入预先训练的显著分类器，得到所述显著分类器输出的所述显著置信度，将所述伪装图像特征输入预先训练的伪装分类器，得到所述伪装分类器输出的所述伪装置信度；
    所述根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，包括：
    利用预先训练的解码器，根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定所述目标识别结果；
    所述显著编码器、所述伪装编码器、所述显著分类器、所述伪装分类器和所述解码器，根据预设的样本图像集联合训练得到，所述样本图像集包括多个显著样本图像和多个伪装样本图像，所述显著样本图像中所述目标对象为显著类型，所述伪装样本图像中所述目标对象为伪装类型。
  5. 根据权利要求4所述的方法，其中，所述显著编码器、所述伪装编码器、所述显著分类器、所述伪装分类器和所述解码器是通过以下方式联合训练得到的：
    获取样本输入集和样本输出集，所述样本输入集包括：多个样本输入，每个所述样本输入包括所述样本图像集中的一个样本图像，所述样本图像为所述显著样本图像或所述伪装样本图像；所述样本输出集中包括与每个所述样本输入对应的样本输出，所述样本输出为对应的所述样本图像的真实识别结果，所述真实识别结果用于标识对应的所述样本图像中目标对象所在的区域；
    将每个所述样本输入作为所述显著编码器的输入，得到显著样本图像特征，并将该样本输入作为所述伪装编码器的输入，得到伪装样本图像特征；
    将所述显著样本图像特征输入所述显著分类器，得到第一置信度，并将所述伪装样本图像特征输入所述伪装分类器，得到第二置信度；
    将目标样本图像特征输入所述解码器，以得到所述解码器输出的样本识别结果，若该样本输入为所述显著样本图像，所述目标样本图像特征为所述显著样本图像特征，若该样本输入为所述伪装样本图像，所述目标样本图像特征为所述伪装样本图像特征；
    将所述样本识别结果和该样本输入对应的样本输出作为判别器的输入，以得到所述判别器输出的判别结果；
    根据所述显著样本图像特征、所述伪装样本图像特征、所述第一置信度、所述第二置信度、所述样本识别结果、所述判别结果确定总损失；
    根据所述总损失联合训练所述显著编码器、所述伪装编码器、所述显著分类器、所述伪装分类器、所述解码器和所述判别器。
  6. 根据权利要求5所述的方法，其中，所述根据所述显著样本图像特征、所述伪装样本图像特征、所述第一置信度、所述第二置信度、所述样本识别结果、所述判别结果确定总损失，包括：
    根据所述显著样本图像特征和所述伪装样本图像特征，确定相似度损失；
    根据所述第一置信度和所述第二置信度确定置信度损失；
    根据所述样本识别结果和每个所述样本输入对应的样本输出，确定识别损失；
    根据所述判别结果和每个所述样本输入对应的样本输出，确定对抗损失；
    根据所述相似度损失、所述置信度损失、所述识别损失、所述对抗损失，确定所述总损失。
  7. 根据权利要求6所述的方法，其中，所述根据所述显著样本图像特征和所述伪装样本图像特征，确定相似度损失，包括：
    根据所述显著样本图像特征和所述伪装样本图像特征的余弦相似度确定所述相似度损失；
    所述根据所述第一置信度和所述第二置信度确定置信度损失，包括：
    将所述第一置信度和第一目标真值的交叉熵损失，与所述第二置信度和第二目标真值的交叉熵损失求和，得到所述置信度损失，所述第一目标真值和所述第二目标真值根据对应的样本输入确定；
    所述根据所述样本识别结果和每个所述样本输入对应的样本输出，确定识别损失，包括：
    将所述样本识别结果与每个所述样本输入对应的样本输出的交叉熵损失，与所述样本识别结果与每个所述样本输入对应的样本输出的Dice损失求和，得到所述识别损失；
    所述根据所述判别结果和每个所述样本输入对应的样本输出，确定对抗损失，包括：
    将所述判别结果和第三目标真值的交叉熵损失，与每个所述样本输入对应的样本输出和第四目标真值的交叉熵损失求和，得到所述对抗损失，所述第三目标真值和所述第四目标真值根据对应的样本输入确定。
  8. 一种组织图像的识别装置，其中，所述装置包括：
    获取模块，用于获取内窥镜采集的组织图像；
    提取模块，用于分别提取所述组织图像的显著图像特征以及伪装图像特征；
    置信度确定模块，用于根据所述显著图像特征确定所述组织图像的显著置信度，并根据所述伪装图像特征确定所述组织图像的伪装置信度；
    识别模块，用于根据所述显著图像特征、所述伪装图像特征、所述显著置信度和所述伪装置信度，对所述组织图像进行识别，以确定目标识别结果，所述目标识别结果用于标识所述组织图像中目标对象所在的区域。
  9. 一种计算机可读介质，其上存储有计算机程序，其中，该程序被处理装置执行时实现权利要求1-7中任一项所述方法的步骤。
  10. 一种电子设备，包括：
    存储装置，其上存储有计算机程序；
    处理装置，用于执行所述存储装置中的所述计算机程序，以实现权利要求1-7中任一项所述方法的步骤。
  11. 一种计算机程序产品，包括计算机程序，所述计算机程序被处理装置执行时实现权利要求1-7中任一项所述方法的步骤。
  12. 一种计算机程序，所述计算机程序被处理装置执行时实现权利要求1-7中任一项所述方法的步骤。
PCT/CN2023/082017 2022-03-30 2023-03-16 组织图像的识别方法、装置、可读介质和电子设备 WO2023185497A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210326583.1A CN114495176B (zh) 2022-03-30 2022-03-30 组织图像的识别方法、装置、可读介质和电子设备
CN202210326583.1 2022-03-30

Publications (1)

Publication Number Publication Date
WO2023185497A1 true WO2023185497A1 (zh) 2023-10-05

Family

ID=81488718

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/082017 WO2023185497A1 (zh) 2022-03-30 2023-03-16 组织图像的识别方法、装置、可读介质和电子设备

Country Status (2)

Country Link
CN (1) CN114495176B (zh)
WO (1) WO2023185497A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114495176B (zh) * 2022-03-30 2022-12-06 北京字节跳动网络技术有限公司 组织图像的识别方法、装置、可读介质和电子设备

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200193269A1 (en) * 2018-12-18 2020-06-18 Samsung Electronics Co., Ltd. Recognizer, object recognition method, learning apparatus, and learning method for domain adaptation
CN111931844A (zh) * 2020-08-07 2020-11-13 商汤集团有限公司 图像处理方法及装置、电子设备和存储介质
CN114495176A (zh) * 2022-03-30 2022-05-13 北京字节跳动网络技术有限公司 组织图像的识别方法、装置、可读介质和电子设备

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8526723B2 (en) * 2009-06-23 2013-09-03 Los Alamos National Security, Llc System and method for the detection of anomalies in an image
CN111898538B (zh) * 2020-07-30 2024-04-12 平安科技（深圳）有限公司 证件鉴伪方法、装置、电子设备及存储介质
CN112633384B (zh) * 2020-12-25 2022-11-01 北京百度网讯科技有限公司 基于图像识别模型的对象识别方法、装置和电子设备
CN112801169B (zh) * 2021-01-25 2024-02-06 中国人民解放军陆军工程大学 一种基于改进yolo算法的伪装目标检测方法、系统、装置及存储介质
CN113536978B (zh) * 2021-06-28 2023-08-18 杭州电子科技大学 一种基于显著性的伪装目标检测方法
CN113449727A (zh) * 2021-07-19 2021-09-28 中国电子科技集团公司第二十八研究所 一种基于深度神经网络的伪装目标检测识别方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200193269A1 (en) * 2018-12-18 2020-06-18 Samsung Electronics Co., Ltd. Recognizer, object recognition method, learning apparatus, and learning method for domain adaptation
CN111931844A (zh) * 2020-08-07 2020-11-13 商汤集团有限公司 图像处理方法及装置、电子设备和存储介质
CN114495176A (zh) * 2022-03-30 2022-05-13 北京字节跳动网络技术有限公司 组织图像的识别方法、装置、可读介质和电子设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
AIXUAN LI; JING ZHANG; YUNQIU LV; BOWEN LIU; TONG ZHANG; YUCHAO DAI: "Uncertainty-aware Joint Salient Object and Camouflaged Object Detection", ARXIV.ORG, 6 April 2021 (2021-04-06), XP081933485 *

Also Published As

Publication number Publication date
CN114495176B (zh) 2022-12-06
CN114495176A (zh) 2022-05-13

Similar Documents

Publication Publication Date Title
CN111476309B (zh) 图像处理方法、模型训练方法、装置、设备及可读介质
CN111368685B (zh) 关键点的识别方法、装置、可读介质和电子设备
WO2023061080A1 (zh) 组织图像的识别方法、装置、可读介质和电子设备
US20210334604A1 (en) Facial recognition method and apparatus
WO2022252881A1 (zh) 图像处理方法、装置、可读介质和电子设备
WO2023030373A1 (zh) 组织腔体的定位方法、装置、可读介质和电子设备
WO2023207564A1 (zh) 基于图像识别的内窥镜进退镜时间确定方法及装置
WO2023185497A1 (zh) 组织图像的识别方法、装置、可读介质和电子设备
WO2023185516A1 (zh) 图像识别模型的训练方法、识别方法、装置、介质和设备
WO2023030427A1 (zh) 生成模型的训练方法、息肉识别方法、装置、介质及设备
WO2023138558A1 (zh) 一种图像场景分割方法、装置、设备及存储介质
CN114863124A (zh) 模型训练方法、息肉检测方法、相应装置、介质及设备
CN114332033A (zh) 基于人工智能的内窥镜图像处理方法、装置、介质及设备
CN111783674A (zh) 一种基于ar眼镜的人脸识别方法和系统
CN112287945A (zh) 碎屏确定方法、装置、计算机设备及计算机可读存储介质
WO2023165332A1 (zh) 组织腔体的定位方法、装置、可读介质和电子设备
CN116152576B (zh) 图像处理方法、装置、设备及存储介质
CN110349108B (zh) 处理图像的方法、装置、电子设备、及存储介质
CN114937178B (zh) 基于多模态的图像分类方法、装置、可读介质和电子设备
EP4447006A1 (en) Font recognition method and apparatus, readable medium, and electronic device
CN110765304A (zh) 图像处理方法、装置、电子设备及计算机可读介质
CN114240843B (zh) 图像检测方法、装置、计算机可读存储介质以及电子设备
CN116434287A (zh) 一种人脸图像检测方法、装置、电子设备及存储介质
WO2022052889A1 (zh) 图像识别方法、装置、电子设备和计算机可读介质
CN114610938A (zh) 遥感图像检索方法、装置、电子设备和计算机可读介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23777872

Country of ref document: EP

Kind code of ref document: A1