CN110046666B

CN110046666B - Mass picture labeling method

Info

Publication number: CN110046666B
Application number: CN201910312598.0A
Authority: CN
Inventors: 何志权; 许琦; 何志海
Original assignee: Shenzhen Deepvision Creative Technology Ltd
Current assignee: Shenzhen Deepvision Creative Technology Ltd
Priority date: 2019-04-18
Filing date: 2019-04-18
Publication date: 2022-12-02
Anticipated expiration: 2039-04-18
Also published as: CN110046666A

Abstract

The invention provides a method for labeling a mass of pictures, which comprises the following steps: step 1, establishing an initial model according to labeling information of a plurality of pictures; step 2, predicting the unmarked picture by using the initial model; step 3, carrying out confidence coefficient analysis on the prediction result in the step 3, thereby selecting reliable prediction; step 4, replacing the prediction result of reliable prediction with labeling information; and 5, updating the initial model by using the marking information obtained in the step 4, and returning to the step 1 to repeat iteration. The method can quickly and effectively label massive pictures so as to solve the problem of the real bottleneck of data labeling in deep learning.

Description

Mass picture labeling method

Technical Field

The invention relates to the technical field of picture marking, in particular to a massive picture marking method.

Background

Deep learning is currently finding increasingly widespread use in both academic and industrial areas. However, the deep learning model has three resource limitation problems: the data size of model training is large, and if the model training depends on network transmission, the network bandwidth becomes a bottleneck; a large amount of computing resources, many deep learning model parameters and complex operation process result in huge computation amount; a large amount of annotation data is required to train complex models.

Annotating data is a tedious process and is typically performed point-by-point with a brush aimed at a target, as shown in fig. 1, where (a) is the original picture and (b) is the annotated picture. Therefore, the existing mode wastes time and labor, and in the labeling process, the labeling effect is often not ideal due to artificial factors and differences.

The mass pictures are marked by consuming a large amount of manpower and material resources, but a method for quickly and effectively marking the mass pictures does not exist at present.

Disclosure of Invention

The invention provides a method for labeling a mass of pictures, which aims to solve at least one technical problem.

In order to solve the above problem, as an aspect of the present invention, a method for labeling a large number of pictures is provided, including: step 1, establishing an initial model according to labeling information of a plurality of pictures; step 2, predicting the unmarked picture by using the initial model; step 3, carrying out confidence coefficient analysis on the prediction result in the step 3, thereby selecting reliable prediction; step 4, replacing the prediction result of reliable prediction with labeling information; and 5, updating the initial model by using the marking information obtained in the step 4, and returning to the step 1 to repeat iteration.

Preferably, the labeling information in step 1 is obtained by manual labeling.

Preferably, step 3 comprises: and setting a confidence coefficient function for evaluating the reliability of the prediction, wherein the function utilizes the original image information and the output vector of the softmax layer of the model to evaluate the prediction result.

Preferably, step 3 comprises: and judging the matching degree of each point in the prediction result by utilizing the pixels in the 4x4 field around the point to calculate the information such as the variance and the gradient of the contrast and the output vector of the softmax layer of the position model, and if the matching degree is matched according to the preset confidence threshold, the prediction is reliable.

Preferably, the anastomotic match is characterized by a neural network or a support vector machine.

Preferably, the updating the initial model by using the label information obtained in step 4 includes: let the initial model be θ _t The last used model is θ' _t-1 Then the new model is θ = α θ' _t-1 +(1-α)θ _t Where α is an exponential smoothing coefficient.

By adopting the technical scheme, the method can quickly and effectively label massive pictures to solve the real bottleneck problem of labeling data in deep learning, firstly labels a small number of pictures, trains a deep learning model, then predicts the remaining unmarked pictures by using the model, evaluates the prediction result, converts accurate prediction into labeling, and then optimizes the model, so that the accuracy of the model is higher and the labeling accuracy is higher and higher in circulating iteration.

Drawings

FIG. 1 is a schematic diagram illustrating a prior art labeling of a picture;

FIG. 2 schematically illustrates a flow chart of the present invention;

fig. 3 schematically shows a schematic diagram of selecting reliable prediction results according to confidence.

Detailed Description

Embodiments of the invention will be described in detail below with reference to the drawings, but the invention can be implemented in many different ways as defined and covered by the claims.

An initial model is assumed to exist, and if not, a small number of pictures can be artificially marked to train the initial model. For convenience of the following explanation, we take defect detection as an example, as shown in fig. 1.

Step 1: the unlabeled data is predicted with the initial model.

Step 2: and carrying out confidence degree analysis on the predicted result. Since the model performance of step 1 is not ideal at this time, the predicted result is good or bad, and therefore, the purpose of performing confidence analysis on the predicted result is to select reliable prediction.

Confidence function

To evaluate the reliability of model predictions. This function evaluates the curve part in (b) of fig. 1 using the original image information I and the output vector of the softmax layer of the model. Specifically, for each point on the curve, we calculate the variance, gradient, etc. of the contrast using the pixels in the 4 × 4 domain around the point, and at this position, the softmax layer output of the model represents the predicted defect strength. These two pieces of information, if matched with a high degree of agreement, indicate that the prediction is reliable. The mapping relationship can be characterized by a simple neural network or a support vector machine (support vector machine).

And step 3: and selecting a key sample. Step 2, confidence analysis is carried out on all predictions, and accordingly, reliable predictions can be selected. We can set a confidence threshold T, when the threshold is greater than T, indicating that the prediction is reliable, and leave, otherwise leave, as shown in fig. 3, the prediction confidence in the upper box in (b) is low, and leave in the lower box.

And 4, step 4: and converting the prediction result into the labeling information. The prediction result with high confidence and the artificial label are consistent on a large probability and can be converted into label information.

And 5: and updating the model. By using more labeled information, we can train out a model theta _t The last used model is θ' _t-1 Then the new model is θ = α θ' _t-1 +(1-α)θ _t Where α is an exponential smoothing coefficient. The model has better prediction performance than the previous model.

With the updated model, we return to step 1 to predict the unmarked part and update the existing prediction. The performance of the model is better and better by repeating the iteration, and the quality of the label is gradually improved.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for labeling massive pictures is characterized by comprising the following steps:

step 1, establishing an initial model according to labeling information of a plurality of pictures, wherein the labeling information is obtained in a manual labeling mode;

step 2, directly utilizing the initial model to predict the unmarked picture;

step 3, performing confidence analysis on the prediction result in the step 2, thereby selecting reliable predictions, which comprises the following steps: setting a confidence function for evaluating the reliability of prediction, wherein the function utilizes original image information and an output vector of a softmax layer of a model to evaluate a prediction result, judging the matching degree of each point in the prediction result between the variance and the gradient of the contrast ratio of each point and the output vector of the softmax layer of the model by utilizing pixels in the 4x4 field around each point, and if the point is matched according to a preset confidence threshold value, the prediction is reliable;

step 4, replacing the prediction result of reliable prediction with labeling information;

and 5, updating the initial model by using the marking information obtained in the step 4, wherein the method comprises the following steps: setting an initial model as

The last used model is

Then the new model is

Wherein

Is an exponential smoothing coefficient; then, the iteration is repeated by returning to the step 1.

2. The method for labeling the massive pictures according to claim 1, wherein the matching is characterized by being characterized by a neural network or a support vector machine.