CN110136098A - A deep learning-based cable sequence detection method - Google Patents

Abstract

The invention discloses a cable sequence detection method based on deep learning. The difference operation is performed on the gray value of the pixel point, a threshold is selected and the difference result is compared, and the positioning result map is output. Step 3: According to the characteristics of the three-color cable image, a simplified feature extraction network, an optimized anchor frame, and an ELU activation function are proposed to improve Faster R‑CNN; Step 4: Use the improved Faster R‑CNN target detection algorithm for line Sequence detection of cables, and statistics of accuracy and detection time. The method of the invention can more fully extract the characteristics of the three-color cable, has the advantages of convenient operation and high efficiency, saves a lot of manpower consumption, and the improved algorithm adopted can simultaneously reduce the cable detection time, reduce false detection, error It can greatly improve the accuracy of cable detection.

Description

A deep learning-based cable sequence detection method

technical field

The invention relates to the field of image processing, and specifically designs a cable sequence detection method based on deep learning.

Background technique

Existing cable sequence testing requires a lot of manual testing. Taking this method will consume a lot of human resources, manual inspection will affect the efficiency of detection, and the cost will be high, and the use of high-intensity manual inspection is likely to cause false detection and false detection, which cannot meet the needs of today's industrial automation production; existing The cable detection algorithm lacks versatility, the detection results are not ideal, and the practicability is not very high, which has a certain impact on the product quality and production efficiency of the enterprise.

SUMMARY OF THE INVENTION

In view of the deficiencies in the above background technology, the present invention provides a fast and accurate deep learning-based cable sequence detection method in order to solve the existing technical problems.

The technical scheme of the present invention is as follows: a deep learning-based cable sequence detection method, comprising the following steps:

Step 1: First, grayscale the three-color cable image, which is the basis for subsequent sequential detection; Step 2: Perform a differential operation on the gray values of adjacent pixels in a certain line of the image, select a threshold and compare the difference results, and output Positioning result map; Step 3: According to the characteristics of the three-color cable image, a simplified feature extraction network, an optimized anchor frame, and an ELU activation function are proposed to improve Faster R-CNN; Step 4: The improved Faster R-CNN is improved The algorithm performs cable sequence detection, and counts the accuracy rate and detection time.

Further, in the step 2, in the grayscale image of the three-color cable area, the grayscale value of a certain row of pixels is taken. The difference operation is performed on the gray value of adjacent pixels in this row.

Further, in the step 2, the threshold is set to 5, the coordinates of the pixels larger than the threshold 5 are counted, and finally the coordinates larger than the threshold 5 are output and the positioning result map of the three-color cable area is output.

Further, in the simplified feature extraction network in step 3, a simplified shared convolutional layer network is used to extract features, and one convolutional layer in the convolutional layer 3 and convolutional layer 4 in the shared convolutional layer network is deleted respectively. .

Further, some optimizations are made to the anchor frame in the said step 3, the designed anchor frame includes the cable area as much as possible, and the optimized anchor frame adopts the anchor frame extraction of 1:1.5, 1:1, 1:2 target area features.

Further, in the step 3, in the Fast R-CNN network, the ELU function is used instead of the ReLU function.

Further, there are 4000 cable images in total, and the cable images are divided into two image sets: 2800 as the training set and 1200 as the test set

1. First, grayscale the cable image, and convert the three-color cable image into a grayscale image;

2. Take the gray value of a certain line of the image and perform a differential operation on the gray values of adjacent pixels;

3. Obtain a suitable threshold m through experiments, compare the result obtained in step 2 with the threshold, output the coordinates greater than the threshold, and output the cable positioning result map.

4. An improved Faster R-CNN target detection algorithm is proposed, an anchor frame that is more in line with the characteristics of cable images and a simplified feature extraction network are selected;

5. The Faster R-CNN target detection algorithm is used to detect the cable sequence. Finally, mAP index is used to evaluate the effect of the model.

The invention has the following beneficial effects: the method of the invention adopts the improved Faster R-CNN target detection algorithm to detect the cable sequence, firstly locates the cable region of interest, and then uses the simplified convolution feature extraction network and optimized anchor frame , ELU activation function to improve the Faster R-CNN algorithm. The method of the invention can more fully extract the characteristics of the three-color cable, has the advantages of convenient operation, high efficiency, and saves a lot of manpower consumption.

The invention deletes the conv3_4 layer and the conv4_4 layer in the conv3 and conv4 in the shared convolutional layer network respectively, and the optimized anchor frame adopts the anchor frame with the aspect ratio of 1:1.5, 1:1, 1:2 to extract the target area feature , when the threshold value m=5, the positioning effect of the three-color cable area is ideal. The cleverness of the improvement is that it can reduce the cable detection time at the same time, and greatly improve the detection speed; in addition, it can also effectively reduce the False detection, false detection and greatly improve the accuracy of cable detection.

For further effects, the convolutional layer and pooling layer of the shared convolutional layer have been improved to a certain extent, and the conv3_4 and conv4_4 layers in conv3 and conv4 in the shared convolutional layer network are respectively deleted, which can also improve the model. The ELU activation function is used to replace the original ReLU activation function. The left side of the function image has soft saturation, which helps to improve the noise robustness. The characteristics of the ELU function will make the output mean close to zero and speed up the convergence. .

For further effects, for the original Faster R-CNN, anchor boxes with aspect ratios of 2:1, 1:1, and 1:2 are used, and the optimized anchor boxes use aspect ratios of 1:1.5, 1:1, and 1:2. The anchor frame extracts the features of the target area, and the anchor frame designed in this way can also more fully extract the features of the image.

Description of drawings

Figure 1 is the network structure diagram of Faster R-CNN algorithm;

Figure 2 is the ReLU function diagram;

Figure 3 shows the shared convolution layer;

Figure 4 is an image of the cables in the correct order.

Figure 5 is an image of cables in the wrong order.

Figure 6 is the grayscaled image.

FIG. 7 is an image after positioning the cable area.

Figure 8 is a schematic diagram of the proposed optimized anchor box.

Fig. 9 is the ELU function diagram;

Figure 10 shows the result of detection using the improved Faster R-CNN algorithm.

Detailed ways

The following briefly introduces the network structure principle of the Faster R-CNN algorithm related to the present invention.

As shown in Figure 1, it is the network structure diagram of Faster R-CNN algorithm. The Faster R-CNN algorithm is mainly composed of two modules: the original image passes through the shared convolution layer and enters two network modules respectively. The first one is the feature layer, ROI layer, and fully connected layer connected in turn in the Fast R-CNN network module. , the final processed image is classified or regressed; the second is the feature map connected in turn in the RPN module of the region proposal network, a low-rank vector, and the final processed image is classified or regressed; the Fast R-CNN network module, the main purpose It is to detect and identify the target category in the candidate region generated by the region generation network; the region proposal network RPN module, the main purpose is to generate the candidate region. The region proposal network is connected to the convolutional layer after the shared convolutional layer, and a sliding window network is used in the last convolutional layer of the RPN network. The position of each sliding window corresponds to the original image, and the corresponding position in the original image That is, the regional proposal boxes for the target, and these regional proposal boxes are called anchor boxes. The anchor box contains three different areas {128*128, 256*256, 512*512}, and the different aspect ratios are {1:1, 1:2, 2:1}.

As shown in Figure 2, it is the ReLU activation function in the Fast R-CNN network module. The ReLU function expression is:

f(x)=max(0,x)

x is the input of the neural network node, and f(x) is the input and output of the neural network node;

As shown in Figure 3, this shared convolutional layer is a model with 16 layers: it has a total of five convolutions, namely convolutional layer 1 to convolutional layer 5, and each convolution is followed by a pooling layer. Convolutional layer 1 and convolutional layer 2 respectively contain 2 convolutional layers, convolutional layer 3 to convolutional layer 5 contain 3 convolutional layers respectively, and fully connected layer 1, fully connected layer 2, fully connected layer 3 . The pooling method is used between the convolutional layers, which is mainly to retain the main features of the extraction, while reducing the parameters and calculation amount of the next layer to prevent overfitting. Finally, the classifier is used for classification.

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

In this embodiment, a method for detecting cable sequence is provided, which includes the following steps:

The cable image set has a total of 4000 images, and the cable images are divided into two image sets: 2800 as the training set and 1200 as the test set.

1. As shown in Figures 4 and 5, Figure 4 is the cable image in the correct order, and Figure 5 is the core wire image in the wrong order. As shown in Figure 6, the color cable image is first converted into a grayscale image, and the grayscale image has less pixel information than the color image, which reduces the amount of calculation;

2. In the grayscale image of the three-color cable area, take the grayscale value of a row of pixels. Perform a differential operation on the adjacent pixel grayscale values of this row to obtain some differential values;

3. When the threshold m=5 is obtained through experiments, the positioning effect of the three-color cable area is ideal, so the threshold is set to 5. Compare this threshold with the difference value obtained in the second step, count the coordinates of the pixels greater than the threshold m, and finally output the coordinates greater than the threshold m and output the positioning result map of the three-color cable area. The result is shown in Figure 7. Show;

4. Some improvements have been made to the previous feature network extraction part in the Faster R-CNN target detection algorithm. According to the uncomplicated characteristics of the three-color cable image, the conditions for the number of network layers for extracting features do not need to be very high. The efficiency of the model, the convolutional layer and the pooling layer of the shared convolutional layer have been improved to a certain extent, and the conv3_3 layer in conv3 and the conv4_3 layer in conv4 in the shared convolutional layer network are deleted respectively, without reducing the accuracy rate. Under the premise, the detection speed is greatly improved;

5. According to the characteristics of the three-color cable image, the anchor frame is optimized, and the designed anchor frame includes the cable area as much as possible, so as to complete the detection task better. The characteristic of the three-color cable image is that the height is more than the length. The analysis shows that the anchor frame with the original aspect ratio of 2:1 cannot properly correspond to the three-color cable image. As shown in Figure 8, the anchor frame with an aspect ratio of 2:1 is changed to 1:1.5, and the anchor frames of 1:1 and 1:2 are retained. On the one hand, the function is to better match the cable of this article. Images, on the other hand, are experimentally verified to be able to save a lot of computer memory.

6. The ELU activation function is used to replace the original ReLU activation function. There is soft saturation on the left side of the function image, which helps to improve noise robustness. The characteristics of the ELU function will make the output mean close to zero and speed up the convergence; among them, ELU The function can activate some neurons. It combines the characteristics of Sigmoid and ReLU. The left side of the function image has soft saturation, which helps to improve noise robustness. The characteristics of the ELU function will make the output mean close to zero, speeding up the convergence.

The ELU function diagram shown in Figure 9, the ELU function expression:

x is the input of the neural network node, f(x) is the input and output of the neural network node, and α is an adjustable parameter;

5. Using Faster R-CNN target detection algorithm for sequential detection. As shown in Figure 10, finally, the mAP index is used to evaluate the effect of the model. In order to reflect the detection effect of the method provided by the present invention, the accuracy and detection time of the method provided by the present invention and the original target detection algorithm are compared. Table 1 shows the accuracy rate and detection time of the two methods. It can be seen from Table 1 that the method provided by the present invention improves a certain detection accuracy, consumes less network memory, and has better detection performance.

Table 1: Cable Test Results

method mAP(%) Detection time (S) Original Faster R-CNN 94.32 0.284 proposed method 96.55 0.236

In the description of this specification, reference to the terms "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples", etc., is meant to incorporate the embodiments A particular feature, structure, material, or characteristic described by an example or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, The scope of the invention is defined by the claims and their equivalents.

Claims (7)

1. a kind of order of cables detection method based on deep learning, which comprises the following steps:

Step 1: gray processing being carried out to three colo(u)r streak cable images first, does basis for subsequent order detection；Step 2: a certain to image Capable neighbor pixel gray value carries out calculus of differences, chooses a threshold value and difference result compares, and exports positioning result figure；Step Rapid 3: the characteristics of being directed to three colo(u)r streak cable images proposes anchor frame, the ELU activation primitive pair of simplified feature extraction network and optimization Faster R-CNN is improved；Step 4: improved Faster R-CNN algorithm being subjected to order of cables detection, and counts accurate Rate and detection time.

2. a kind of order of cables detection method based on deep learning according to claim 1, which is characterized in that described In step 2, in the grayscale image in three colo(u)r streak cable regions, the gray value of certain a line pixel is taken.To the pixel ash that this line is adjacent Angle value carries out calculus of differences.

3. a kind of order of cables detection method based on deep learning according to claim 1, which is characterized in that described In step 2, setting threshold value is 5, counts the coordinate of the pixel greater than threshold value 5, and finally output is greater than the coordinate of threshold value 5 and defeated The positioning result figure in three colo(u)r streak cable regions out.

4. a kind of order of cables detection method based on deep learning according to claim 1, which is characterized in that described The feature extraction network that step 3 simplifies, it is specific that feature is extracted using simplified shared convolutional layer network, it deletes respectively shared A convolutional layer in convolution layer network in convolutional layer 3 and convolutional layer 4.

5. a kind of order of cables detection method based on deep learning according to claim 1, which is characterized in that described Some optimizations are done to anchor frame in step 3, cable region is included in the inside as far as possible by the anchor frame of design, and the anchor frame of optimization uses The anchor frame of 1:1.5,1:1,1:2 extract target area feature.

6. a kind of order of cables detection method based on deep learning according to claim 1, which is characterized in that described Step 3, in FastR-CNN network, ReLU function is replaced using ELU function.

7. a kind of order of cables detection method based on deep learning according to claim 1, which is characterized in that cable figure Piece totally 4000, cable picture is divided into two pictures: 2800 are used as training set, and 1200 are used as test set.