A cloth defect recognition method based on deep convolutional neural network

Abstract

The present invention relates to a method for identifying cloth defects based on a deep convolutional neural network, comprising the following four parts: (1) modifying the Faster R-CNN neural network model, including using a global flat pooling layer to replace the first two fully connected layers in the feature extraction network, and adding anchor boxes in the region proposal network; (2) using a data enhancement method to expand the data set; (3) using a variable learning rate to train the improved Faster R-CNN network model; (4) using an industrial linear array camera to capture cloth images, and using the trained network model to identify defects, and outputting defect categories and location information. The present invention can identify various types of cloth defects such as broken weft, broken warp, holes, floating weft, etc., and has a faster recognition speed and higher recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the application of deep learning technology in the field of machine vision inspection in the textile industry, and specifically to a cloth defect recognition method based on a deep convolutional neural network. Background Art

[0002] China is a textile country, and the textile industry plays an important role in the country's economic development. Among them, cloth quality is a key issue restricting the development of the textile industry. With the rapid progress of science and technology, the competition in the international textile industry is becoming increasingly fierce. Major textile companies are facing tremendous pressure from high quality standards and high labor costs. The price of cloth with defects is often about 50% lower than that of normal cloth. However, at present in China, the cloth defect detection of most textile companies still remains in the traditional manual inspection stage. Under traditional manual inspection, the moving speed of the cloth is usually only 5-10m / min, which leads to low efficiency of the cloth inspection process. Moreover, after a period of highly concentrated work, the human eye will inevitably produce visual fatigue, resulting in many defects that should have been detected being missed. The missed detection rate of cloth defects by human eyes is as high as more than 30%. In recent years, with the rapid development of optical technology, digital circuit technology and image processing technology, machine vision has been more and more widely used in the field of industrial surface inspection, and the automatic detection system of cloth defects has become an inevitable trend.

[0003] For the automatic defect detection system of cloth, the core part is the defect detection method. The existing methods are mainly divided into several categories: based on statistical analysis, frequency domain analysis, model analysis, learning analysis, etc. Histogram statistical technology is used for cloth defect detection due to its low computational cost. Zhang et al. performed defect detection by checking the histogram difference between dyed and defect-free cloth images (see Zhang WY, Zhang J, Hou Y, et al. MWGR: A new method for real-time detection of cord fabric defects. International Journal of Advanced Mechatronic Systems, 2012: 458-461.). Although this method has the characteristics of simplicity and high computational speed, it has the problem of low detection accuracy. Common frequency domain analysis methods include Gabor filtering, wavelet transform, etc. Karlekar et al. proposed a cloth texture detection method combining wavelet transform and image morphology, which achieved good results (see Karlekar VV, Biradar MS, Bhangale K B. Fabric Defect Detection Using Wavelet Filter. IEEE International Conference on Computing Communication Control and Automation, 2015: 712-715).

[0004] Patent application number 201410200849.3 discloses a fabric defect detection and classification method based on the Otsu method. In this method, the lower computer performs filtering, interpolation, variance sampling preprocessing, Otsu method segmentation and binarization, hole filling and small block processing, binary image channel connection processing, extracts and saves each defect area; the upper computer divides the defects into regional and non-regional defects, warp defects and weft defects, dark defects and bright defects. This method has high recognition efficiency and meets real-time requirements, but the identified defect type is relatively single, and the detection effect is poor for cloth with complex textures.

[0005] Since the great success of the deep neural network AlexNet in 2012, excellent classification networks such as VGG-16 and VGG-19 (see Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition [J]. Computer Science, 2014), Inception, and ResNet have emerged one after another. The use of deep convolutional neural networks (CNN) has prompted machine vision research to enter a new stage. At present, the mainstream convolutional neural network target detection and recognition algorithms are: 1) target detection and recognition algorithms based on region proposals, with representative ones being R-CNN, Fast R-CNN, and Faster R-CNN (see Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39 (6): 1137-1149), which cover almost every field from classification to detection, segmentation, and positioning; 2) target detection and recognition algorithms based on regression, with representative ones being YOLO and SSD. Jing Junfeng et al. achieved a detection rate of 87.5% for color fabric defects based on a defect detection algorithm based on a deep convolutional neural network, which achieved a good visualization effect for the defect position and shape of the color fabric (see Jing Junfeng, Fan Xiaoting, Li Pengfei, et al. Color fabric defect detection using deep convolutional neural network [J]. Journal of Textile Research, 2017, 38 (2): 68-74.). However, this method has many neural network parameters and high computational complexity. Moreover, it can only classify defects but cannot locate them. It is necessary to use other methods or manual methods to first detect the defect location and extract the defect image, and then classify the defects. It cannot be directly applied to actual detection tasks. Patent application number 201810052130.8 discloses a fabric defect detection method based on deep learning. This method divides the images in the sample library into a set of image blocks; establishes a six-layer deep neural network, and adjusts the connection weights of the fifth and sixth layers using the BP algorithm to output the type corresponding to the defect; divides the image to be detected into image blocks and inputs them into the neural network; counts the detection results of all image blocks contained in the detection image, and gives the quality grade of the cloth based on this, showing the specific coordinate position of the defect. Although this method uses a convolutional neural network with fewer layers, the detection speed is improved, but in actual situations, there are many types of defects and complex features, and the recognition accuracy is insufficient. Summary of the invention

[0006] In order to automatically detect and classify defects in grey cloth, bleached cloth, dyed cloth and yarn-dyed cloth, the present invention provides a cloth defect recognition method based on deep convolutional neural network, which specifically includes the following four parts: Faster R-CNN neural network modification, cloth defect data set preparation, neural network training optimization, and real-time cloth defect detection and classification.

[0007] (1) Faster R-CNN neural network modification

[0008] The deep convolutional neural network used in the present invention is modified based on Faster R-CNN, specifically including the following two aspects:

[0009] (1-1) Faster R-CNN feature extraction network modification

[0010] The detection of cloth defects must have both high detection accuracy and good real-time effect. Combining these two points, this method finally chooses the VGG-16 network as the feature extraction network and optimizes the VGG-16 network for cloth defect classification.

[0011] The VGG-16 network uses continuous 3×3 convolution kernels to replace larger convolution kernels, which can improve the depth of the network while ensuring the same receptive field, thereby improving the network's feature extraction capabilities. The VGG-16 network is specifically composed of 13 convolutional layers and 3 fully connected layers: 3×3 filters are used in the convolution layer, and every 2 or 3 filters are stacked continuously to form a convolution sequence to imitate a larger receptive field effect; 2×2 maximum pooling is used in the pooling layer with a step size of 2 for dimensionality reduction operations after convolution; the fully connected layer is composed of 3 fully connected combinations with 4096, 4096, and 1000 channels respectively; finally, it is input to the SoftMax classifier for classification output.

[0012] Although the structure of VGG-16 is very simple, it consumes a lot of computing resources. Most of the parameters come from the first fully connected layer. The first fully connected layer FC1 has 4096 nodes, and the previous pooling output has 7×7×512=25088 nodes. The fully connected layer requires 4096×25088 nodes. This fully connected layer alone accounts for 74% of the number of parameters, which consumes a lot of computing resources.

[0013] A higher detection speed is required in the cloth detection process, but too many parameters in the fully connected layer will greatly slow down the detection speed. In order to reduce the complexity of defect detection of cloth images, the present invention adopts a global average pooling layer (GlobalAverage Pooling, GAP) to replace the fully connected layer, which can not only reduce the number of parameters and speed up network operations, but also can well reduce the occurrence of overfitting. The global average pooling layer directly uses the mean of the feature map to represent the confidence map of a class. For example, if there are 10 classes, 10 feature maps are output at the end. All the feature values ​​in each feature map are added together to find an average value. The 10 average values ​​obtained are the probability or confidence of the corresponding class. Since the number of parameters of the global average pooling layer is zero, it not only reduces the parameters to speed up the calculation speed and prevent data redundancy, but also avoids the risk of overfitting due to too many parameters in the fully connected layer. At the same time, the characteristics of the black box operation of the fully connected layer are also eliminated, and the actual category meaning of each channel is directly given. The present invention specifically replaces the first two fully connected layers FC1 and FC2 in the VGG-16 network with a global average pooling layer, and the third fully connected layer FC3 remains unchanged. The input to the global average pooling layer is a 7×7×512 feature map, and each channel is subjected to a 7×7 average pooling operation. After global average pooling, 512 feature maps are generated, and the feature values ​​represent the channel features. The feature values ​​of each channel are then input into the original third fully connected layer FC3 to further purify the features, and the features further purified by the fully connected layer are classified and output. Using a global average pooling layer to replace the first two fully connected layers reduces the number of parameters by more than 75%, greatly reduces resource usage, and optimizes the running speed.

[0014] (1-2) Add Faster R-CNN anchor box

[0015] The core concept of the Faster R-CNN network model is an end-to-end design mode, which requires the location and identification of the target area in the scene to be output simultaneously. The innovative work lies in the proposal of Region Proposal Networks (RPN). RPN is a fully convolutional neural network (FCN), whose input is a feature map of any size and output is a series of rectangular target boxes. To generate candidate regions, a small network is used to make sliding window selections on the output of the last layer of the shared convolutional layer. The input of the network is an n×n window of the feature map. For each pixel, k target candidate regions are predicted at the same time. These target candidate regions are associated with this window and are called anchor boxes.

[0016] In the original Faster R-CNN network, two parameters are set for generating anchor boxes, namely ratios = [0.5, 1, 2] representing the aspect ratio condition for generating anchor boxes, and scales = [8, 16, 32] representing the area condition for generating anchor boxes. After convolution feature extraction, the original image becomes 1 / 16 of the original input image. The sizes of 8×8, 16×16, and 32×32 in scales correspond to the anchor box sizes of the original image, which are (16×8)×(16×8)=128×128, (16×16)×(16×16)=256×256, and (16×32)×(16×32)=512×512 areas, respectively. Each area of ​​each size is transformed according to the aspect ratio of 1:2, 1:1, and 2:1 in ratios. For example, the generated forms of 16×16 size are 23×12, 16×16, and 11×22, and the pixel area remains unchanged. Combining all transformations, 9 anchor boxes are generated for each window, and then each candidate box is evaluated.

[0017] If the anchor boxes are not adjusted correctly, the neural network will never know the existence of certain small, large or irregular objects and will never have a chance to detect them. Increasing the diversity of anchor boxes helps with multi-scale defect detection. Increasing the types of anchor boxes, making ratios = [0.25, 0.5, 1, 2], scales = [4, 8, 16, 32], each window can generate 16 scale-varying anchor boxes. The improved network is more accurate for multi-scale defect detection and small target detection.

[0018] (2) Fabric defect dataset preparation

[0019] Since the amount of cloth defect image data is relatively small, the frequency of defects in the whole roll of cloth is very low, the actual collection is difficult, and there is no unified standard public data set. Therefore, after collecting basic cloth defect images through the camera, the data set can be expanded by image flipping, scaling, and fusion to achieve data enhancement, and the defects are divided into seven categories: repeated warp and weft, broken warp and weft, reed marks, spots, holes, creases, and stains. Image fusion is shown in formula (1), ImageM is the fused image, Image1 and Image2 are two cloth defect images with the same resolution; α is the weight coefficient of Image1, and its value range is [0,1], β is the weight coefficient of Image2, and its value range is [0,1], and α+β=1; the form of image fusion includes inter-class combination and intra-class combination.

[0020] ImageM=α×Image1+β×Image2 (1)

[0021] (3) Neural network training optimization

[0022] During the training process, the learning rate is an important hyperparameter in deep learning. It determines whether the objective function can converge to the local minimum and appropriately converge to the minimum. When the learning rate is set too small, the convergence process will become very slow. When the learning rate is set too large, the gradient may oscillate back and forth near the minimum value, or even fail to converge. When the learning rate is fixed, when it reaches the convergence state, it will swing in a larger area near the optimal value; and when the learning rate is reduced as the number of iterations increases, it will swing in a smaller area near the optimal value when converging. Therefore, choosing a suitable learning rate is crucial for model training.

[0023] In order to achieve rapid convergence in the initial stage of training and stabilize to the local optimal value in the later stage, the present invention adopts the improved Faster R-CNN network model in (1) with a variable learning rate training: the learning rate lr is initially set to 0.001, so that the model can converge quickly; when the total number of training iterations reaches n1, the learning rate lr is reduced to 0.0001; when the total number of training iterations reaches n2, the learning rate lr is reduced to 0.00001, so that the objective function can stably converge to the local minimum, until the iteration is completed to obtain the trained model parameters; wherein the value range of n1 is [10000, 500000], the range of the number n2 is [20000, 1000000], and n2 is required to be greater than n1. The larger the number of samples in the data set, the larger the values ​​of n1 and n2.

[0024] (4) Real-time detection and classification of cloth images

[0025] (4-1) Use an industrial line array camera to capture a fresh cloth image.

[0026] (4-2) Use the trained improved Faster R-CNN neural network to detect defects in cloth images and obtain the location and type of defects in the image. The network first uses 13 shared convolutional layers of VGG-16 to extract cloth image features and generate feature maps, which are then passed through the RPN network to generate regional scores and regional proposals; then the feature map is convolved with a dimension of 256, a kernel size of 3×3, and a step size of 1. According to the RPN network, each anchor box is classified and learned to distinguish whether it belongs to the foreground or the background, and the classification box position is regressed and fine-tuned to the appropriate position, and the classification score and the position coordinates of each box after regression are output; then the ROI area is generated according to the regional opinion of RPN, and each ROI area is unified to a size of 7×7. Each ROI area contains 512 channels. The ROI area is mapped to the sample space through the global average pooling layer and the fully connected layer, and the classification score in each sample space and the regressed target frame are obtained; finally, the detected defect category and position information are output according to the classification score and the regressed target frame position information.

[0027] (4-3) Repeat steps (4-1) to (4-2) to complete the inspection of the entire roll of cloth.

[0028] The technical concept of the present invention is as follows: a global average pooling layer is used to replace the first two fully connected layers in the Faster R-CNN feature extraction network to achieve the purpose of reducing network parameters and accelerating network operation speed; an anchor point frame of the RPN network in Faster R-CNN is increased to improve the network's detection effect on defect detection of various scales and small target detection; during the training process, a variable learning rate is used to enable model training to converge quickly and to converge stably to a local minimum; the model parameters obtained by training are used to perform defect detection on the original cloth image collected by an industrial linear array camera to achieve better detection positioning and defect classification effects.

[0029] Compared with the existing technology, the method has the following beneficial effects: it supports the detection and classification of various types of cloth defects including broken weft, broken warp, holes, floating weft, etc., reduces the number of network parameters by improving Faster R-CNN, improves the detection effect, and can accurately detect defects in images with complex texture information and various color backgrounds. It can achieve high-efficiency, high-accuracy and fast detection, improves the efficiency of cloth detection in industrial sites, and reduces the impact of false detection caused by human factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of the present invention;

[0031] Figure 2The defect image data enhancement effect diagram in the present invention, wherein the right image of (a) is the image flip effect diagram of the left image, the right image of (b) is the image zoom effect diagram of the left image, and the right image of (c) is the image fusion effect diagram of the two left images;

[0032] Figure 3 The figure is a comparison between the VGG-16 network structure before and after modification in the present invention and the prior art, wherein the upper solid line path is the modified network structure, and the lower dotted line path is the network structure of the prior art;

[0033] Figure 4 The figure is a comparison diagram of the detection effect after adding the anchor point frame in the present invention, wherein the upper figure is a detection effect diagram using the original anchor point frame, and the lower figure is a detection effect diagram after adding the anchor point frame;

[0034] Figure 5 This is a diagram showing the effect of cloth defect recognition using the improved Faster R-CNN in the present invention. DETAILED DESCRIPTION

[0035] The present invention will be described in detail below with reference to the embodiments and the accompanying drawings, but the present invention is not limited thereto.

[0036] The fabric used in the defect recognition embodiment of the present invention is dyed fabric. The processing platform selected by the present invention is a combination of Inteli9-9900k, NVIDIA RTX2080ti and 32G RAM, and the operating system is LINUX64Ubuntu16.04. The deep learning frameworks include Caffe, Pytorch, TensorFlow, etc. Considering that TensorFlow has good visualization effects and Faster R-CNN is easy to configure on TensorFlow, the method of the present invention is finally implemented on the TensorFlow framework.

[0037] like Figure 1 The cloth defect recognition method based on deep convolutional neural network includes four parts:

[0038] (1) Faster R-CNN neural network modification;

[0039] (2) Fabric defect data set preparation;

[0040] (3) Neural network training optimization;

[0041] (4) Real-time detection and classification of cloth images.

[0042] The first part of the Faster R-CNN neural network modification specifically includes:

[0043] (1-1) Fully connected layer replacement

[0044] The first two fully connected layers FC1 and FC2 in the original network are replaced by a global average pooling layer (Global AveragePooling, GAP), and the third fully connected layer FC3 remains unchanged. The input to the global average pooling fully connected layer is a 7×7×512 feature map. After global average pooling, 512 feature maps are generated, and the features are further purified by the fully connected layer before classification and output. Using a global average pooling layer to replace the first two fully connected layers reduces the number of parameters by more than 75% in total. The detection speed of Faster R-CNN after the improvement using the global average pooling layer is increased by nearly 40% compared with before the improvement. The test experiment was carried out on the experimental platform of the present invention. It took only 0.051s to detect an image with a resolution of 640×480, and the detection accuracy was not much different from that before the improvement.

[0045] (1-2) Add Faster R-CNN anchor box

[0046] On the basis of the original network, we increase the types of anchor boxes, making ratios = [0.25, 0.5, 1, 2], scales = [4, 8, 16, 32], which represents four anchor boxes of pixel area size, namely 4×4, 8×8, 16×16, 32×32. Each pixel area is generated in the form of aspect ratios of 1:4, 1:2, 1:1, and 2:1. Each window can generate 16 anchor boxes of varying scales, such as Figure 5 The improved network is more accurate for defect detection of various scales and small objects. By adjusting the anchor boxes, the average precision (mAP) of Faster R-CNN for all class labels on the experimental data set in the second part reached 0.8935, which is a significant improvement over the 0.7448 before adjustment.

[0047] The second part of the cloth defect data set production specifically includes:

[0048] Since the amount of cloth defect image data is relatively small, the frequency of defects in the entire roll of cloth is very low, the actual collection is difficult, and there is no public data set with unified standards. Therefore, 520 basic defect images of dyed cloth were collected by camera and then the images were flipped as follows: Figure 2 (a) shows that the image is scaled as Figure 2 As shown in (b), image fusion is as follows Figure 2(c) The data set is expanded to achieve data enhancement. The original 520 cloth defect data set is expanded to 2080, and the defects are divided into seven categories: repeated warp and weft, broken warp and weft, reed marks, spots, holes, creases, and stains. Image fusion is shown in formula (1), ImageM is the fused image, Image1 and Image2 are two cloth defect images with the same resolution; α is the weight coefficient of Image1, the value range is [0,1], and it is 0.5 here; β is the weight coefficient of Image2, the value range is [0,1], and it is 0.5 here, satisfying α+β=1; the form of image fusion includes inter-class combination and intra-class combination.

[0049] ImageM=α×Image1+β×Image2 (1)

[0050] The third part of neural network training optimization specifically includes:

[0051] In the present invention, as the training progresses, the learning rate is gradually reduced. At the beginning of the training, the learning rate lr is set to 0.001 so that the model can converge quickly; when the total number of training iterations reaches n1, the learning rate lr is reduced to 0.0001; when the total number of training iterations reaches n2, the learning rate lr is reduced to 0.00001 so that the objective function can stably converge to the local minimum; the value range of the number n1 is [10000, 500000], here it is 40000, the range of the number n2 is [20000, 1000000], here it is 60000, and the total number of iterations of a complete training is 80000. The mAP is further improved to 0.9237 by adjusting the learning rate during the training process.

[0052] The fourth part of real-time detection and classification of cloth images specifically includes:

[0053] (4-1) Use an industrial camera to capture the original image of the cloth.

[0054] (4-2) Use the trained improved Faster R-CNN neural network to detect defects in cloth images and obtain the location and type of defects in the image. The network first uses 13 shared convolutional layers of VGG-16 to extract cloth image features and generate feature maps, which are then passed through the RPN network to generate regional scores and regional proposals; then the feature map is convolved with a dimension of 256, a kernel size of 3×3, and a step size of 1. According to the RPN network, each anchor box is classified and learned to distinguish whether it belongs to the foreground or the background, and the classification box position is regressed and fine-tuned to the appropriate position, and the classification score and the position coordinates of each box after regression are output; then the ROI area is generated according to the regional opinion of RPN, and each ROI area is unified to a size of 7×7. Each ROI area contains 512 channels. The ROI area is mapped to the sample space through the global average pooling layer and the fully connected layer, and the classification score in each sample space and the regressed target frame are obtained; finally, the detected defect category and position information are output according to the classification score and the regressed target frame position information.

[0055] (4-3) Repeat steps (4-1) to (4-2) to complete the inspection of the entire roll of cloth.

Claims

1. A cloth defect recognition method based on deep convolutional neural network, characterized in that: The identification method comprises the following steps: (1) Faster R-CNN neural network modification, including the following sub-steps: (1-1) The first two fully connected layers FC1 and FC2 in the feature extraction network VGG-16 of Faster R-CNN are replaced by a global average pooling layer, and the third fully connected layer FC3 remains unchanged; (1-2) Based on the 9 predicted anchor boxes of the original region generation network of Faster R-CNN, the scale size and scale change types of the anchor boxes are increased to generate more than 9 anchor boxes with different scales; (2) Collect the original data of cloth defects and use data enhancement methods to expand the data set; (3) Neural network training optimization: Use a variable learning rate to train the improved Faster R-CNN network model in (1); (4) Real-time detection and classification of cloth images: (4-1) Use an industrial camera to capture the original image of the cloth; (4-2) Use the trained improved Faster R-CNN neural network to detect defects in the cloth image and obtain the location and type of the defects in the image, which includes the following sub-steps: (4-2-1) Use 13 shared convolutional layers of VGG-16 to extract cloth image features and generate feature maps; (4-2-2) Perform convolution on the feature map with a dimension of 256, a convolution kernel size of 3×3, and a stride of 1. Perform classification learning on each anchor box based on the RPN network to distinguish whether it belongs to the foreground or the background, and regress the classification box position to the area where the defect is located, and output the classification score and the position coordinates of each box after regression; (4-2-3) Generate ROI regions based on the regional opinions of RPN, and unify the ROI regions to 7×7 size. Each ROI region contains 512 channels. The ROI region passes through the global average pooling layer and the fully connected layer to map the features to the sample space, and obtain the classification score in each sample space and the regressed target border; (4-2-4) According to the classification score and the position information of the target frame after regression, the classification score of each target frame is sorted and the class with the highest score is obtained. When the highest score exceeds the set threshold TH, the target frame is the defect class with the highest score. Otherwise, there is no defect in the target frame. The detected defect category and position information are output. The value range of TH is [0,1]; (4-3) Repeat steps (4-1) to (4-2) to complete the inspection of the entire roll of cloth.

2. The method for identifying cloth defects based on a deep convolutional neural network according to claim 1, characterized in that: In the step (1-2), 16 anchor boxes of varying scales are generated, the aspect ratio conditions of the anchor boxes are generated as ratios=[0.25, 0.5, 1, 2], and the area conditions of the anchor boxes are generated as scales=[4, 8, 16, 32].

3. The cloth defect recognition method based on deep convolutional neural network according to claim 1, characterized in that: The step (2) includes the following sub-steps: (2-1) Collecting basic fabric defect images through a camera; (2-2) Expand the data set by image flipping, image scaling, and image fusion to achieve data enhancement; (2-3) Classify the defects.

4. The method for identifying cloth defects based on a deep convolutional neural network according to claim 3, characterized in that: In the step (2-2), the image fusion is as shown in formula (1): ImageM=α×Image1+β×Image2 (1) Wherein, ImageM is the fused image, Image1 and Image2 are any two cloth defect images with the same resolution and image format, α is the weight coefficient of Image1, and its value range is [0,1], β is the weight coefficient of Image2, and its value range is [0,1], α+β=1.

5. The method for identifying cloth defects based on a deep convolutional neural network according to claim 3, characterized in that: In the step (2-3), the defects are classified into seven categories: repeated warp and weft, broken warp and weft, reed marks, spots, holes, folds, and stains. The image fusion includes inter-class combination and intra-class combination of defects.

6. The method for identifying cloth defects based on a deep convolutional neural network according to claim 1, characterized in that: The step (3) includes the following sub-steps: (3-1) Set the initial learning rate lr to 0.001, and train the improved Faster R-CNN network model in step (1) with the data set in step (2); (3-2) When the total number of training iterations reaches n1, the learning rate lr will be reduced to 0.0001, where the value range of n1 is [10000, 500000]; (3-3) When the total number of training iterations reaches n2, the learning rate is reduced to 0.00001 until it converges to the local optimal iteration and the trained model parameters are obtained. The number of iterations n2 is in the range of [20000, 1000000] and n2 is required to be greater than n1.

Images