An experimental operation platform unlabeled object detection method

Abstract

The application discloses an experimental operation platform unlabeled object detection method, comprising the following steps: processing an original picture to obtain a labeled segmented picture; training the labeled segmented picture and the original picture; processing a picture with an object to be detected to realize unlabeled object detection. The application realizes rapid detection of different objects, and can replace the object to be detected, thereby reducing detection cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to an experimental operation platform unlabeled object detection method. BACKGROUND

[0002] With the rise of deep learning, computer vision technology has been continuously developed, among which the object detection technology represented by target detection and instance segmentation is particularly remarkable. It has good applications in various industries. In student experimental operation evaluation projects, the relative positions of experimental equipment can be used to judge the correctness of student operation. Compared with target detection, instance segmentation can more accurately reflect the relative positions of equipment, thereby better assisting judges to make judgments, such as determining the density of solid objects, exploring the density of solids, and other experiments with greater openness.

[0003] However, the existing detection technology needs a large number of labeled objects to achieve the purpose of detection when applied, and the measured objects cannot be replaced during actual experimental operation, resulting in high detection cost. SUMMARY

[0004] The purpose of the present application is to provide an experimental operation platform unlabeled object detection method to realize rapid detection of different objects and reduce detection cost.

[0005] To solve the above technical problems, the present application provides an experimental operation platform unlabeled object detection method, which comprises the following steps:

[0006] Processing the original picture to obtain a labeled segmentation picture;

[0007] Training the labeled segmentation picture and the original picture;

[0008] Processing the picture with the object to be detected to realize the detection of unlabeled objects.

[0009] Further, the processing step of the original picture comprises:

[0010] Labeling the original picture;

[0011] Training the instance segmentation network to obtain a segmentation picture based on the original picture.

[0012] Further, the training step of the segmentation picture and the original picture is:

[0013] Training the labeled segmentation picture and the original picture through the training model, so that the training model can restore the original picture according to the labeled segmentation picture.

[0014] Further, the training model is a picture translation model.

[0015] Further, the processing step of the picture with the object to be detected includes:

[0016] Infer the picture with the object to be detected through the segmentation network, and the object to be detected is labeled with an error label;

[0017] Restore the labeled segmentation picture through the picture translation model to obtain a reconstructed picture;

[0018] Differentially compare the reconstructed picture with the original picture to realize detection of the object to be detected.

[0019] Further, the specific steps of the differential comparison include:

[0020] Process the reconstructed picture and the original picture through a coding and decoding network to obtain a score map;

[0021] Analyze the score map to confirm the region where the object to be detected is located.

[0022] Further, the coding and decoding network is a U-shaped coding and decoding network based on a twin neural network.

[0023] Further, the processing steps of the reconstructed picture and the original picture include:

[0024] Encode the reconstructed picture and the original picture through the coding and decoding network, extract features, and share weights;

[0025] Simultaneously calculate and forwardly transmit the features of the reconstructed picture and the original picture;

[0026] Return a difference value using deconvolution of the decoder.

[0027] Further, the features of the reconstructed picture and the original picture are fused into the decoding network after being extracted.

[0028] Further, the analysis step of the score map includes:

[0029] Mark the difference values in the score map in the reconstructed picture in the form of scattered points, and the region with a larger value is the region where the object to be detected is located.

[0030] Compared with the prior art, the present application has at least the following beneficial effects: an unannotated object is labeled with an error label by using a trained instance segmentation network, a segmentation picture is restored through a trained image-to-image translation model, thereby generating a difference with an original picture, a U-shaped coding and decoding network based on a twin neural network is used to compare the difference between the restored segmentation picture and the original picture, thereby determining the position of the unannotated object, and detection of the unannotated object is realized; the unannotated object can be replaced, and rapid detection is realized. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 A flowchart of a detection method in an embodiment of the present application;

[0032] Figure 2 A flowchart of a score map generation in an embodiment of the present application;

[0033] Figure 3 A specific structure of a coding and decoding network in an embodiment of the present application. DETAILED DESCRIPTION

[0034] The experimental operation platform unlabeled object detection method of the present application will be described in more detail below with reference to the accompanying drawings, in which a preferred embodiment of the present application is shown. It should be understood that those skilled in the art can modify the present application described herein while still achieving the advantageous effects of the present application. Therefore, the following description should be understood as a broad knowledge to those skilled in the art, and not as a limitation of the present application.

[0035] The present application will be described in more detail in the following paragraphs with reference to the accompanying drawings. The advantages and features of the present application will be more apparent from the following description and claims. It should be noted that the drawings are very simplified and use non-precise proportions, only to facilitate, clearly assist in the purpose of illustrating the embodiments of the present application.

[0036] As shown in Figure 1 An experimental operation platform unlabeled object detection method is provided in an embodiment of the present application, comprising the following steps:

[0037] Processing the original picture to obtain a labeled segmentation picture;

[0038] Training using the labeled segmentation picture and the original picture;

[0039] Processing the picture with the object to be detected to achieve the detection of the unlabeled object. The segmentation picture is obtained by processing the original picture, and the model is trained to quickly detect the object to be detected; and the object to be detected can be replaced to improve the detection efficiency.

[0040] The following preferred embodiments of the experimental operation platform unlabeled object detection method are listed to clearly illustrate the content of the present application. It should be clear that the content of the present application is not limited to the following embodiments, and other improvements by conventional technical means of those skilled in the art are also within the scope of the idea of the present application.

[0041] Please refer to Figure 2 In an embodiment, the processing step of the original picture comprises:

[0042] annotating the original picture;

[0043] training an instance segmentation network to obtain a segmented picture based on the original picture. The segmentation network first reduces the resolution of the original picture to improve feature expression capability, and then increases the resolution of the original picture to improve the ability to distinguish details.

[0044] In an embodiment, the training steps of the segmented picture and the original picture are as follows:

[0045] The annotated segmented picture and the original picture are trained through the training model, so that the training model can restore the original picture according to the annotated segmented picture. The segmented picture is input into the training model, so that the training model outputs a picture, and through deep learning, the picture output by the training model is close to the original picture.

[0046] Preferably, the training model is a picture translation model. The training model is a pix2pix model, which is an image-to-image translation based on a generative adversarial network (GAN). Unlike other image translation methods, which are based on a random noise generated image, the pix2pix model learns the mapping between the input image and the output image as a condition, so as to obtain a specified output image.

[0047] The generative adversarial network is a deep learning model, and is one of the most promising methods for unsupervised learning on complex distribution in recent years. The model learns through the mutual game of at least two modules in the framework: the generation model and the discrimination model to produce a very good output.

[0048] In an embodiment, the processing steps of the picture with the object to be detected include:

[0049] The picture with the object to be detected is processed through the segmentation network, and the object to be detected is labeled with an error label.

[0050] The annotated segmented picture is restored through the picture translation model to obtain a reconstructed picture.

[0051] The reconstructed picture and the original picture are compared to realize detection of the object to be detected. By comparing the reconstructed picture and the original picture, the difference is the position of the object to be detected. By detecting the difference, the object to be detected is detected.

[0052] Preferably, the specific steps of the difference comparison include:

[0053] The reconstructed picture and the original picture are processed through a coding and decoding network to obtain a score map.

[0054] The score map is analyzed to determine the region where the object to be detected is located. The coding and decoding network first reduces the resolution of the original picture through the encoder to improve the feature expression capability, then improves the resolution of the original picture through the decoder to improve the detail distinguishing capability, extracts the feature difference between the reconstructed picture and the original picture, obtains the score map, and determines the region where the object to be detected is located through the difference on the score map.

[0055] Preferably, the coding and decoding network is a U-shaped coding and decoding network designed based on a twin neural network. The twin neural network is connected through the shared weights of the network, and when the similarity of two pictures needs to be compared, the comparison is realized by comparing the picture features. The neural network can better extract the features of the related pictures. In order to ensure that the features extracted from the two pictures to be compared are in the same domain, the neural network used for extraction needs to share a feature weight. Therefore, the twin neural network can effectively compare the differences between the two pictures.

[0056] Please refer to Figure 3 Preferably, the processing steps of the reconstructed picture and the original picture are:

[0057] The coding and decoding network is used to code the reconstructed picture and the original picture, extract features and share weights;

[0058] The features of the reconstructed picture and the original picture are calculated and forward propagated at the same time;

[0059] The deconvolution of the decoder is used to return the difference value. The encoder gradually reduces the resolution of the original picture, extracts features and shares weights, calculates and forward propagates the features of the two pictures at the same time, the decoder gradually improves the resolution of the original picture, improves the detail distinguishing capability, returns the difference value through the deconvolution of the decoder, and realizes the difference comparison between the restored reconstructed picture and the original picture.

[0060] Preferably, the features of the reconstructed picture and the original picture are fused into the decoding network at the same time after being extracted. The features extracted at each stage are fused into the decoding network through the connection of two or more arrays (Concat), so as to avoid the loss of features in the decoding process.

[0061] Preferably, the analysis step of the score map is:

[0062] The difference value in the score map is marked in the reconstructed picture in the form of a scatter plot, wherein the region with a larger value is the region where the object to be detected is located. The variables in the data set usually have different dimensions and orders of magnitude. When the values of different variables differ greatly, if the original values are directly used for analysis, the effect of the variable with a larger value will be highlighted, and the effect of the variable with a smaller value will be weakened. In order to eliminate the influence of dimensions and orders of magnitude on data analysis, the data needs to be standardized. The score map is a commonly used data standardization method. The score map is to mark all standardized values in the form of a scatter plot in the picture, so that the vertical comparison of variables with different orders of magnitude and dimensions becomes possible, and the distribution of the standardized values of all variables can be previewed. For classification data, the distinguishing ability of variables for different category samples can also be overviewed.

[0063] In summary, the unannotated object is labeled with an error label by using the trained instance segmentation network, the segmentation picture is restored by using the trained image-to-image translation model, and the difference from the original picture is generated. The U-shaped encoding and decoding network designed based on the twin neural network is used to compare the difference between the restored segmentation picture and the original picture, so as to determine the position of the unannotated object, and the detection of the unannotated object is realized. The unannotated object can be replaced, and rapid detection is realized.

[0064] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for detecting unlabeled objects on an experimental operating platform, characterized in that, The method includes the following steps: The original image is processed to obtain an labeled segmented image. The processing steps for the original image include: Label the original image; Train the instance segmentation network to obtain segmented images based on the original images; The training process utilizes labeled segmented images and the original images. The training steps for the segmented images and the original images are as follows: The labeled segmented images and the original images are trained together using a training model, which enables the training model to reconstruct the original images from the labeled segmented images. The training model is an image translation model. The image containing the object to be detected is processed, and the processing steps of the image containing the object to be detected include: The image containing the object to be detected is passed through the segmentation network for inference, and the object to be detected is incorrectly labeled. The labeled segmented images are restored using the image translation model to obtain reconstructed images; The reconstructed image is compared with the original image to detect the object to be detected.

2. The experimental operation platform described in claim 1 does not specify an object detection method, characterized in that, The specific steps for the difference comparison include: The reconstructed image and the original image are processed by an encoding and decoding network to obtain a score image; Analyze the score map to identify the area where the object to be detected is located.

3. The experimental operation platform described in claim 2 does not specify an object detection method, characterized in that, The encoding / decoding network is a U-shaped encoding / decoding network designed based on a Siamese neural network.

4. The experimental operation platform described in claim 2 does not specify an object detection method, characterized in that, The processing steps for the reconstructed image and the original image are as follows: The reconstructed image and the original image are encoded using an encoding / decoding network, features are extracted and weights are shared; this allows the features of the reconstructed image and the original image to be calculated simultaneously and passed forward. The deconvolution of the decoder returns the difference value.

5. The experimental operation platform described in claim 4 does not specify an object detection method, characterized in that, The features of the reconstructed image and the original image are extracted and then simultaneously fused into the decoding network.

6. The experimental operation platform described in claim 2 does not specify an object detection method, characterized in that, The analysis steps for the score graph are as follows: The difference values ​​in the score image are marked as scatter points in the reconstructed image, with the areas with larger values ​​being the areas where the objects to be detected are located.

Images