49 results about "Class activation mapping" patented technology

The invention belongs to the field of computer vision and machine learning, in particular to a method for classifying benign and malignant pulmonary nodules and locating lesions based on 2D depth neural network and multi-slice CT images. The invention uses the multi-slice CT images as the input of the neural network to enable the network to learn the characteristics of different scales and different shapes of the same node, thereby enhancing the robustness and generalization ability of the model. Secondly, it solves the problem that the class activation mapping can not accurately locate the fine-grained region of interest in the image, The invention realizes fine-grained positioning by adding full-link layer to all feature maps of the last layer of the network, even if the mutual influencebetween the final features is weakened, and then combining the feature vector obtained by the full-link layer with a feature vector similar to the feature vector obtained by the U. The combination oflow-dimensional features in the Net structure enables more accurate localization of malignant lesions and improves the accuracy of classification.

The invention provides a target classification and positioning method based on network supervision. The target classification and positioning method comprises the following steps: automatically obtaining a large amount of network image data from a search engine according to the category of a to-be-tested target; filtering to remove noise images to form a training sample set; preliminarily constructing a classification and positioning network; and inputting samples in the training sample set into the preliminarily constructed classification and positioning network to perform feature extraction,classifying the features, obtaining position information of the target object, and training the classification and positioning network. According to the end-to-end fine classification and positioningmethod based on network supervision, massive network images easy to obtain are used as a training set, manual annotation is completely removed, only image-level labels are used, an efficient convolutional network is designed, and algorithms such as global average pooling and class activation mapping graphs are fused, so that the performance of the method exceeds that of a weak supervised learningmethod on fine classification tasks and positioning tasks.

The invention relates to the field of traffic monitoring, in particular to an urban aggregation event prediction and positioning method and device. According to the method and the device, feature extraction is carried out on multiple frames of pictures by utilizing a deep neural network; the time sequence features and the spatial features of multiple frames of pictures are extracted, probability prediction and class activation mapping are carried out by utilizing the time sequence features and the spatial features of the multiple frames of pictures, and the probability and the position of an urban aggregation event are judged, so that the aggregation event can be detected, the urban aggregation event can be predicted in advance, and an aggregation site is positioned. In the process of feature extraction and classification judgment, an ingenious network structure is selected according to actual conditions. According to the method, the urban aggregation event is predicted and positionedin advance based on the deep neural network under the complex Internet of Things mobile data, and long-time-span prediction and positioning can be carried out on a key area where the aggregation eventoccurs or is simulated.

The invention provides a weak-monitoring target positioning method based on data enhancement. The method mainly comprises the steps of reference network construction, target positioning and performance optimization, wherein the steps particularly comprises that for an input picture, firstly a pre-activated residual network is used to achieve an effect of a classification network to serve as a reference network, then a network dataset is used to train the classification network, and positioning performance is optimized through data enhancement, small batching scale and deep network depth, thena class activation mapping (CAM) algorithm is used to generate a heat map, and the reference network outputs results of classification (object tag) and positioning (bounding box) by controlling a threshold of the heat map. The weak-monitoring target positioning method based on the data enhancement has the advantages that the problem that a departed target positioning method only focuses on a mostdiscriminating part of a target object is solved, a weakly-discriminative part of the target object can be classified and located, and the accuracy of the weak-monitoring target positioning technologycan be improved.

The invention provides a heavy landing analysis method based on a multi-branch time convolutional network. The method comprises the following steps: acquiring original parameter data and a dynamic time point; performing convolution operation on the original parameter data by using the improved time convolution network to generate a feature map of each parameter; performing feature extraction on the feature map to generate overall feature representation; learning a preset category by using the overall feature representation to obtain a parameter level of the preset category and a weight occupied by a feature map of each parameter; and according to the parameter level and the weight occupied by the feature map of each parameter, carrying out linear combination on the feature maps in the overall feature representation to obtain a final class activation mapping map, and according to the class activation mapping map, carrying out analysis on airplane heavy landing. According to the method, a new thought is provided for safety accidents or overrun events in the aviation field, reference is provided for interpretability work of the time sequence classification problem, technical reference is provided for flight safety, and the method has good theoretical and application values.

The invention discloses a behavior recognition method, system and device based on a human skeleton and a medium. The method comprises the following steps: acquiring spatio-temporal sequence data of ahuman skeleton; inputting the space-time sequence data into a behavior recognition model to obtain a behavior recognition result; wherein the behavior recognition model comprises a first processing unit, a class activation mapping unit, a second processing unit and a classification unit; wherein the first processing unit is connected to the second processing unit through the class activation mapping unit, and the first processing unit and the second processing unit are connected to the classification unit. According to the recognition method provided by the embodiment of the invention, the twodata streams are processed through the first processing unit and the second processing unit, and richer and more discriminative features can be extracted for behavior recognition, so that behaviors and actions of a human body can be recognized more accurately and effectively, and the recognition accuracy is improved. The method can be widely applied to the technical field of computer vision.

The invention discloses an ultrahigh-pixel tissue pathological image segmentation method, and belongs to the field of image processing and artificial intelligence. The method comprises the following steps of: S1, randomly selecting pathological section image blocks with fixed window sizes on an ultrahigh-pixel tissue pathological image on which tissue pathological labeling is completed to form a pathological image block training data set; S2, preprocessing the pathological section image blocks; s3, establishing a multi-scale space full convolution network and a class activation mapping model thereof, and training network parameters of the model by adopting a pathological image block training set in combination with a full connection layer so as to realize accurate lesion classification and recognition based on image blocks; and S4, inputting the to-be-analyzed ultra-high pixel tissue pathological image by using the multi-scale space full convolution network structure, and outputting an image slice segmentation result with pathological tissue position information. The ultra-high pixel tissue pathological image segmentation method can efficiently and accurately realize pixel-level accurate region segmentation of the ultra-high pixel tissue pathological image.

The invention discloses a medical image classification method and device, and relates to the technical field of digital image processing. The method comprises the steps: migrating a deep learning model trained in a natural image data set to the field of medical images through transfer learning, and obtaining a basic model of medical image classification; constructing a deep attention branch network based on class activation mapping on the basis of the basic model of medical image classification; constructing a loss weighting module based on a gray level co-occurrence matrix on the basis of the basic model of medical image classification; establishing fusion of a loss weighting module based on a gray level co-occurrence matrix and a deep attention branch network based on class activation mapping on the basis model of medical image classification to obtain a fused medical image classification model; training the fused medical image classification model to obtain a medical image automatic classification model; and automatically classifying the to-be-classified images through the medical image automatic classification model. The method can be used for accurate classification of medical images.

The invention discloses a weak supervision target detection method and a weak supervision target detection system, which are used for training a target detector to detect a target in a picture under the condition of only annotation of an image category, and can save a large amount of manpower, material resources and financial resources consumed by annotation information. In the prior frame generation part, a selective search algorithm and a gradient weighted class activation mapping method are combined to generate a better prior frame, and meanwhile, in the optimization iteration process of a detector, supervision information of low-level features is added, and the concept of likelihood is introduced to measure the degree that a target in the prior frame is a complete target. The problem that a current weak supervision target detection method is prone to falling into a local optimal pain point, so that a network tends to select a prior frame covering a whole target under the condition that no target bounding box information supervision exists is solved. The network improves the performance of weak supervision target detection, and can be used in the fields of automatic driving, intelligent security and protection and the like; experimental results show that the method has good competitive performance.

The invention provides a method for automatically generating a brain CT medical report based on weak supervised attention, relates to the three fields of medical images, computer vision and natural language processing, and designs a weak supervised attention mechanism WGAM to clearly guide an attention model to focus on a focus region so as to improve the accuracy of medical report generation. WGAM is a hierarchical structure and comprises two attention mechanisms of space attention and sequence attention, and the space attention is weakly supervised by a gradient weighted class activation mapping algorithm to obtain a better attention effect. A keyword-driven interactive loop network KIRN is designed as a language generation module to generate a brain CT medical report, a hidden layer state of the language generation module is activated through keyword information containing focus position information, and the accuracy of generating a brain CT image report is improved through dynamic interaction of LSTMword and LSTMsen. According to the method, the work of automatically generating the brain CT medical report is explored for the first time, and effectiveness is achieved.

The invention belongs to the technical field of radar signal modulation mode identification, and particularly relates to a radar signal identification and positioning method based on MobileNet model transfer learning. According to the method, based on MobileNet model transfer learning and gradient weighted class activation mapping, the network model is built by using depth separable convolution, so that model parameters can be effectively reduced, and the calculation efficiency of the model is improved; transfer learning training is carried out by loading the pre-training model in the training process, so that the convergence speed and generalization performance of the model can be improved; and meanwhile, a prediction result of the network model is visualized by adopting a gradient weighted category activation mapping method, so that the interpretability and transparency of the deep learning model are improved.

The invention discloses a deep neural network feature visualization method for constraint optimization class activation mapping. The method comprises the steps of obtaining a pre-training model whichis used for image classification and is constructed by adopting a deep neural network through training or downloading; carrying out forward transmission on a to-be-detected image by using the pre-training model to obtain a feature map, and carrying out further processing to obtain a final weight vector; and performing weighted summation on each component of the feature map through the final weightvector to obtain a visual feature map, and presenting the visual feature map as a final visual result. According to the method, feature visualization can be carried out on any deep neural network, abetter deep feature interpretability visualization effect can be achieved, and the method has less noise and stronger class discrimination.

The invention provides an image emotion classification method based on class activation mapping and visual saliency, and relates to the technical field of computer vision and image processing. The method comprises the following steps: firstly, overall features of an image are extracted through a deep convolutional neural network; saliency detection is carried out on an image by using a multi-scalefull convolutional neural network to further obtain saliency region features of the image, and meanwhile, an emotion distribution diagram of the image is generated through class activation mapping and emotion region features are extracted only by using an emotion label of an image level. The saliency region features and the emotion region features of the image are regarded as local representations of the image, and are further fused with the overall features of the image to obtain more discriminative visual features which are used for visual emotion classification. According to the method, the overall information of the image is considered, the information of the important local area in the image is fully utilized, and meanwhile, only the picture-level emotion label is needed, so that thelabeling burden is greatly reduced.

The embodiment of the invention provides an image retrieval method and device, and the method comprises the steps: employing a convolutional neural network to extract the features of a target image, and enabling the total score of pixels in the feature map at the same position to serve as a class activation mapping table of the target image according to the feature map of the target image output by the last convolutional layer in the convolutional neural network; Multiplying each feature map by the class activation mapping table, performing summation pooling, multiplying the summation poolingresult by the weight of each feature map of the target image, and obtaining intermediate features of each feature map of the target image; Obtaining spatial semantic features of each feature map of the target image according to the target image category probability output by the discrimination layer of the convolutional neural network and the intermediate features of each feature map of the targetimage; And obtaining a retrieval result according to the spatial semantic features of the feature maps of the target image and the pre-obtained spatial semantic features of the feature maps of the to-be-retrieved images. According to the embodiment of the invention, the spatial semantic information is used for retrieval, so that the retrieval precision is improved.

The invention provides a man-machine asynchronous recognition method based on multi-task learning and class activation graph feedback. The method comprises the following steps: training a deep learning model through the multi-task learning in a parameter hard sharing manner, and obtaining the visual interpretation of the trained deep learning model for an output result through a class activation mapping manner; meanwhile, setting a feature region according to the activation domain of the class activation graph of each man-machine asynchronous type; inputting the actually collected original breathing signal into the trained deep learning model, and obtaining an identification result of the current breathing signal; and finally, correcting an identification result according to a feature region set according to a man-machine asynchronous type. Only one network model is trained in a mode that multiple tasks are combined with parameter hard sharing, so that multiple recognized man-machine asynchronous types can be output at the same time through one-time forward calculation, and the recognition efficiency of an existing method is improved. And based on a self-correction classification result fed back by the class activation graph, the method has high accuracy and interpretability.

The invention relates to a training method of a style conversion model, and a training method of a virtual building detection model, and belongs to the technical field of image processing. The training method of a style conversion model comprises the following steps: inputting each sample virtual image and each sample real image into a first generator of a style conversion model to generate a conversion image corresponding to each sample virtual image and an attention thermodynamic diagram of each sample virtual image and each sample real image; inputting each converted image and each real sample image into a discriminator of the style conversion model to generate a discrimination result and an attention thermodynamic diagram of each converted image and each real sample image; respectively generating an adversarial loss value and a class activation mapping loss value according to the discrimination result of the discriminator and each attention thermodynamic diagram; training the style conversion model according to the sum of the adversarial loss value and the class activation mapping loss value. Therefore, the effect of virtualizing virtual images of a model is improved, and the trueness of the virtual images is improved.

The invention discloses a gating feature attention isovariant segmentation method based on weak supervised learning, and the method specifically comprises the steps: 1, training a first classification network, and carrying out the weight sharing, and obtaining a second classification network; training a partial fusion module of the first gating, and carrying out weight sharing to obtain a partial fusion module of the second gating; 2, performing affine transformation on the original image to obtain an affine image; 3, respectively inputting the original image and the affine image into two classification networks; 4, taking the feature layer of the last layer of the two classification networks as class activation mapping and affine class activation mapping; 5, inputting the feature maps output by the two classification networks at the specific stage into a corresponding gated partial fusion module to obtain a gated feature map and an affine gated feature map; 6, inputting results obtained in the step 4 and the step 5 into a cross feature attention model to obtain improved class activation mapping; and 7, realizing image segmentation according to the improved class activation mapping. According to the invention, the segmentation precision of the weak supervision network is improved.

The invention discloses a weak supervision building segmentation method taking a reliable region as attention mechanism supervision, and the method comprises the following steps: constructing a weak supervision semantic segmentation network which comprises a first classification network and a reliable region synthesis module, a second classification network, a pixel attention module, a class activation mapping calculation module, a twin network structure and a loss function design module; building images and manually-marked classification labels are obtained to serve as a training set, the training set is used for training the classification network to obtain initial seeds, and the initial seeds are input into a reliable region synthesis module to obtain reliable labels; training a class activation mapping module based on the pixel attention module and the twin network structure by using the training set to obtain class activation mapping; and finally, taking the generated reliable label as supervision of class activation mapping to obtain a pseudo label, and training an existing network by using the pseudo label to obtain a final building segmentation result. According to the method, pixel-level semantic segmentation is realized only through classification labels.

The invention discloses a general adversarial disturbance generation method based on correlation class activation mapping, and belongs to the field of adversarial machine learning. At present, the key technical problem in the field is deep neural network decision interpretability and adversarial sample mobility enhancement. According to the method, the general adversarial disturbance is generated and optimized by utilizing an inter-layer correlation propagation and class activation mapping cascading mode, and then the focus of the deep neural network is understood. Firstly, a deep neural network classifier is utilized to calculate an original label class and other error label classes of a clean sample, then through forward propagation class activation mapping feature map and correlation coefficient linear weight combination, the contribution of a final thermodynamic diagram of the original label is minimum, the contribution of thermodynamic diagrams of other error classes is maximum, and finally, the error label class is obtained. And the general adversarial disturbance is iteratively updated by minimizing the correlation class activation mapping loss function, so that the general adversarial disturbance with strong mobility is formed, and the attack success rate of the adversarial sample is improved.

The invention discloses a regularization method and a regularization system based on class activation mapping graph guidance. The method comprises the following steps: S1, generating a channel weightfactor and a class activation mapping graph based on a label class by using global pooling layer and full connection layer parameters of a deep neural network; S2, sorting the generated channel weightfactors and the class activation mapping graphs according to contribution degrees of the channel weight factors and the class activation mapping graphs to all channels and space regions of each layerof feature graphs in the neural network; s3, obtaining a feature channel set and a feature point set according to the step S2, and further obtaining binary mask graphs M(1) and M(2) based on a channel weight factor and a class activation mapping graph; s4, generating a random seed binary image M(3) based on Bernoulli distribution, performing logical operation with M(1) and M(2) to obtain a finalbinary mask image M, and thus obtaining a regularized mask image M1; and S5, iteratively carrying out training processes from S1 to S4 for multiple times to complete a regularization optimization process.

The invention discloses an ultrasonic-assisted liver lump diagnosis system and method. The ultrasonic-assisted liver lump diagnosis system comprises an image preprocessing module, a sub-image segmentation module, an RGB channel separation module and a CNN network model construction module. The invention adopts a medical image anomaly detection method integrating the patch-based CNN, the class activation mapping and a faster R-CNN technology, and the method comprises the following steps of firstly, preprocessing an original image to remove the irrelevance; segmenting the original image into a plurality of sub-images, constructing a multi-layer CNN network model for each sub-image, and inputting the sub-images for recognition; connecting the networks of the sub-images as the final model identification result. Compared with the problems of inaccurate subjective judgment and low efficiency caused by the manual scanning judgment in the prior art, a liver abnormal lump image in an ultrasonic image can be automatically recognized, the lumps can be automatically identified for the benign and malignant analysis, and the method and the system are more accurate and efficient.

The invention discloses an adversarial defense method based on class activation mapping. The adversarial defense method comprises the following steps: S1, establishing a comparison image set for maximizing a neuron activation value in a prediction model through a gradient rising method; S2, positioning a judgment region based on the class activation mapping graph of the to-be-detected image; S3, calculating the inconsistency between the to-be-detected image judgment area and the same label comparison image judgment area based on a binarization algorithm; s4, judging whether the to-be-detectedimage has disturbance or not, if the inconsistency degree is greater than a threshold value, determining that the to-be-detected image has countermeasure disturbance, and otherwise, determining that the to-be-detected image is a normal image; and S5, removing the countermeasure disturbance in the to-be-detected image. The adversarial defense method is high in universality, can resist different confrontation attacks, is low in data processing cost, and improves the defense efficiency.