Fluorescence penetrant testing is a non-destructive testing method based on capillary phenomenon to detect surface defects of workpieces [30]. The fluorescent penetrant inspection process for automobile steering knuckles is shown in Figure 1. First, the fluorescent penetrant was applied to the surface of the cleaned automobile steering knuckle and left for 10 min. Through capillary action, the penetrant penetrated into the small cracks on the surface of the workpiece. Then, the excess penetrant was removed by washing with water, and a developer was applied on the surface of the part after drying. Subsequently, the penetrant in the defect was re-adsorbed to the surface of the workpiece under capillary action, thereby forming an enlarged defect display. Finally, by illuminating the steering knuckle with a black light in a dark environment, the cracks on the surface of the steering knuckle could be clearly observed.

The collected fluorescence penetrant inspection images also contained a part of the interference display in addition to the display of crack defect formation. As shown In Figures 1A,C, certain dust residues settled on the workpiece surface and edge, and the geometric mutation of the groove caused the accumulation of permeate to form an interference display. As the defect observation must be conducted under the irradiation of ultraviolet light and owing to the workpiece not being flat, in case of an oblique shooting angle, the collected image contains ultraviolet light reflection stripes, as shown in Figure 1B. In addition, the residual fluorescent penetrant on the surface of the workpiece also interferes with the judgment of crack defects, as shown in Figure 1D.

To analyze the characteristics of the interference display and crack image that can easily occur in the fluorescence penetrant inspection image, the gray distribution map corresponding to the interference display and the crack image in Figure 1 was made, as shown in Figure 2. The dust residues were observed as the workpiece was not cleaned, causing a large amount of permeate to accumulate at its edges. Most of the images were circular and the gray value distribution was between 200–250. The grayscale distribution of the ultraviolet light reflection area transitioned from 100 to 250, and the peak value was similar to the crack grayscale distribution. Further, the edge disturbances appeared as noise points distributed at intervals, and were similar in shape to the cracks. The residual area of the permeate was a fluorescent green display with unstable morphology; however, its gray value was lower than that at the crack. Table 1 presents the gray value and morphological features of the five regional images. Based on the above analysis, we suppressed and removed certain amount of interference through certain image preprocessing methods, and the remaining areas that cannot be removed were input to convolutional neural network (CNN) for identification.

This study developed a fluorescence penetration image acquisition platform comprising a camera obscura device, a purple light lamp, a Daheng industrial camera MER-1220-9GC, and a computer, as shown in Figure 3. According to the process flow of fluorescent penetrant detection, first, the steering knuckle was cleaned with water and dried. This was followed by the even application of a fluorescent penetrant on the surface of the workpiece, which was allowed to stand for a certain period of time. After the effect of the fluorescent penetrant was complete, the excess fluorescent penetrant on the surface of the workpiece was washed off with water applied at low water pressure, and finally the developer was sprayed and dried. After completing the pre-processing of the steering knuckle, it was into the darkroom and turn on the black light. Owing to the large size of the steering knuckle and uneven surfaces, a single camera cannot obtain all the information. Therefore, three industrial cameras are used to examine the steering knuckle from different angles. A total of 56 fluorescent penetrant detection images were captured and 286 images were obtained from related companies; thus, a total of 342 images were collected.

To extract the crack candidate regions and input them into CNN identification, the acquired images must be preprocessed. Based on the feature analysis of the image, the obtained original image was converted into a grayscale image and a median filter was used. The median filter can effectively suppress the intermittent interference points on the edge of the workpiece and certain small noise points in the fluorescence penetrant inspection image, as shown in Figure 5C. As the gray value at the crack was significantly higher than the image background, a global threshold segmentation was performed on the image. Only cracks, ultraviolet light reflections, and certain dust interference areas remained in the segmented image, and certain small noise points and areas with low gray values in the image were removed, as shown in Figure 5D. The finally extracted connected domain corresponded to the window input CNN recognition in the original image.

Data augmentation methods can be divided into two categories based on the method of generating new samples: non-generative data augmentation methods and generative data augmentation methods. Non-generative data enhancement methods mainly include geometric transformation, color gamut transformation, noise injection, random erasing [31], etc. These methods can simulate the position, illumination change, and noise interference in the real situation to a certain extent. The diversity of the dataset, but not total exploitation of the intrinsic characteristics of the original samples, leads to certain limitations of the trained model. In this study, non-generative methods such as random cropping and flipping were used to expand the samples to compare the effect of DCGAN data enhancement.

DCGAN is a combination of convolutional neural network (CNN) and GAN, which improves the quality of generated images and the speed of model convergence. Similar to GAN, DCGAN includes a generator G (Generator) and a discriminator D (Discriminator). Figure 6 shows the DCGAN network structure used in this study. The generator G initially receives a random noise, and then learns the data distribution of the real image to generate a fake image. Thereafter, the discriminator D determined whether the input image is a real image and then achieves the best state through continuous learning. Thus, it checks if the quality of the generated image is sufficiently high such that the discriminator G cannot distinguish whether it is a real or fake image.

The network structure of generator G has a total of five layers, including four layers of deconvolution layers and one layer of fully connected layers. The input was a 100-dimensional random noise that was uniformly distributed. First, a fully connected layer realized the linear transformation of the noise, and then reshaped it into a 4 × 4 × 1,024 three-dimensional matrix. After four layers, the kernel size was 4 × 4, and the step size was 4 × 4. A deconvolution layer of two yields a 64 × 64×3 image. Compared with the GAN network, DCGAN replaces the pooling layer with deconvolution of the corresponding stride. Further, it uses batch normalization Batchnorm (BN layer) at each layer except the output layer to normalize the input of each unit to improve learning stability. Thereafter, the output and other layers uses Tanh and ReLu as the activation functions, respectively, to make the model learn faster to saturate and cover the color space of the training distribution.

The network structure of the discriminator D is similar to the convolutional neural network, and the input is the generated fake crack image. After the fully connected layer, a probability of judging the authenticity of the crack image is obtained. Similar to that in the generator, batch normalization operations are used in the discriminator. In addition, the last layer of the discriminator uses the sigmoid function, and the other layers use the LeakyReLU function to prevent the gradient from disappearing.

The loss function of the DCGAN network parameter update process is expressed as follows: m i n m a x D V D , G = E x ∼ p d a t a x l o g D x + E z ∼ p z z log 1 − D G z (1) where x is the real sample, z is random noise, D x represents the probability that the discriminator judges the real image as real, and G z is the generated image. For generators, the goal is to make D ( G z = 1 , which minimizes E z ∼ p z z log 1 − D G z . For the discriminator, when the input is x , its purpose is to make D x = 1 , which maximizes E x ∼ p d a t a x l o g D x ; whereas, when the input is G z , its purpose is to make D ( G z = 0 , which maximizes E z ∼ p z z log 1 − D G z . During training, the generator and discriminator are trained alternately. First, G is fixed to train D, the parameters in D are updated, followed by the updation of the parameters in G. This process is repeated until the Nash equilibrium state is reached. At this time, the discriminator cannot distinguish the real image of the input image, the output probability stabilizes at 0.5, and the training is complete.

Among the 342 collected images, 200 of them were gridded and the crack areas were selected with a size of 64 × 64 pixels which formed 281 crack area samples. 190 of the 281 samples were used for image generation and CNN model training. The remaining 91 test sets were used for CNN model training. The remaining 142 of the 342 collected images served as the final test for the crack identification algorithm.

The training parameters were set as follows: Batch Size of 32, learning rate set to 0.0002, and the number of training iterations as 1,000; in addition, the Adam optimizer was used.

We selected four generated pictures under different epochs. It is evident that with the increase of the number of training epochs, the quality of the generated cracks continuously increased, as shown in Figure 7. In the first few epochs, the model generated similar noise of the sample. At 150 epochs, the gridding degree of the image disappeared. Further, at 300 epochs, the outline of the crack was observed, and at 500 epochs, the crack was formed with low image clarity. Finally, the crack image generated at 600 epochs was found to be very realistic.The original and the generated crack images were both similar and different. The crack outline, size, background, etc. experienced large or minimal changes. Moreover, the generated image exhibited an independent and identical distribution relationship with the original image, such that the generated crack image was considered suitable as a sample for training the classification network.

After many experiments, it was found that the model collapsed after 1,000 iterations, and the loss function oscillated greatly. This may be because of the data set being exceedingly small and the discriminator was too well trained; thus, the generator could not update the gradient. Therefore, we used crack images generated by the model in 600–1,000 epochs range to expand the dataset.