Single Water Drop Images Download HOT!
Erminia Mckissack
The lens for monitoring the rubber conveyor belt is easy to adhere to a large number of water droplets, which seriously affects the image quality and then affects the effect of fault monitoring. In this paper, a new method for detecting and removing water droplets on rubber conveyor belts based on the attentive generative adversarial network is proposed to solve this problem. First, the water droplet image of the rubber conveyor belt is input into the generative network composed of a cyclic visual attentive network and an autoencoder with skip connections, and an image of removing water droplets and an attention map for detecting the position of the water droplet are generated. Then, the generated image of removing water droplets is evaluated by the attentive discriminant network to assess the local consistency of the water droplet recovery area. In order to better learn the water droplet regions and the surrounding structures during the training, the image morphology is added to the precise water droplet regions. A dewatered rubber conveyor belt image is generated by increasing the number of circular visual attention network layers and the number of skip connection layers of the autoencoder. Finally, a large number of comparative experiments prove the effectiveness of the water droplet image removal algorithm proposed in this paper, which outperforms of Convolutional Neural Network (CNN), Discriminative Sparse Coding (DSC), Layer Prior (LP), and Attention Generative Adversarial Network (ATTGAN).
Rubber conveyor belts [1] have been widely used in coal, mining, port, and other fields, mainly for the transportation of bulk, granular, and powdery solid materials. For a variety of reasons, rubber conveyor belts often show longitudinal tears [2] during operation, causing economic losses and even casualties. With the development of technology, the longitudinal tear detection system of the rubber conveyor belt based on machine vision has been gradually popularized. However, the rubber conveyor belt is prone to dust during transportation of some materials (such as coal and powder ore), which causes the camera lens to become dirty. In order to keep the lenses clean, it is common to spray water on the lenses and then wipe off water. This will leave water droplets on the lens. In addition, in order to reduce the environmental pollution caused by dust, many applications will spray water on the materials. For example, when coal is transported in a coal mine, a large amount of coal dust is in the air. Spraying water is required to reduce the coal dust, which makes the monitoring of water droplets in the lens more common. Therefore, how to effectively remove the water droplets on the rubber conveyor belt monitoring image to ensure the sharpness of the image is an important issue to be solved.
The removal of image water droplets in a rubber conveyor belt is similar to the removal of raindrops in a natural image [3]. At present, the methods of water droplet removal at home and abroad are mainly divided into three categories: filter-based rain removal algorithms [4], sparse coding dictionary- and classifier-based rain removal algorithms [5], and deep learning-based rain removal algorithms [3].
In 2014, Eigen et al. [7] proposed a method for raindrop removing in a single image and trained a convolutional neural network with pairs of raindrop-degraded images and corresponding raindrop-free images to better affect relatively thin and small areas of raindrops or dust. But for larger and dense raindrops, it does not produce good results. In 2015, Luo et al. [5] used the discriminant dictionary to learn the sparse coding method, which improved the accuracy of background layer and rain layer separation. However, when the image contains image texture similar to the rain streak, the image details will be blurred. In 2016, Li [8] proposed single-image rain streak removal using layer prior algorithm based on the Gaussian mixture model, but it is easy to lead to smooth transition in nonrain areas. In 2018, Qian et al. [9] used a network of attentive generative adversarial training raindrop image, and visual attention was given into the generative and discriminative network, which had a good effect on raindrop removal, but there was a defect in losing image detail information. For the rubber belt longitudinal tear monitoring system, under the action of auxiliary light source illumination, the rubber conveyor belt image will have a certain amount of specular reflection effect, and the attention generation is not ideal for the network removal effect.
In this paper, we propose a new method for detecting and removing water droplets from rubber conveyor belts based on the attentive generative adversarial network [9]. The expected results are achieved.
Because the shape of the water droplets on the rubber conveyor belt images is different, the number is different, and the background information occluded by the water droplets is similar to the water droplets so that when the general algorithm detects the water droplets in the image, the area similar to the shape of the water droplet is mistakenly treated as a water droplet. The subsequent operation of removing water droplets will remove the background information in the image, and a clear, water-free background image cannot be restored. Even more difficult is that even if the position of the water drop area is correctly detected, it is impossible to restore a clear water-free background. Therefore, in our method, we utilize a GAN [10] as the backbone of our network, which is recently popular in dealing with the image inpainting or completion problem. Then, our main idea is to inject visual attention [9] into both the generative and discriminative networks. By increasing data preprocessing, improving network structure, and rationally designing network optimizer and hyperparameters, a new method for detecting and removing water droplets in the rubber conveyor belt based on attention generation against the network is designed. A block diagram of the water droplet detection and removal method for the rubber conveyor belt image based on the attentive generative adversarial network is shown in Figure 1.
Firstly, the rubber conveyor belt water droplet image datasets are subjected to data preprocessing, including input normalization, image cropping, image flipping, and image morphology, and then the attention map is added in the generator to make the generation network focus on the area with water droplets and utilize the three loss (perceptual loss, multiscale losses, and attention map loss) functions in [9], and the independent design of the network optimizer make the generation of network training more stable and generate a clear image of the dewater droplets. The generated dewater droplet rubber conveyor belt image is input into the attentive discriminator together with the true clear background image to judge the true and false area of the water drop, and the optimizer and loss function which are most suitable for the discriminator are designed. The image data preprocessing and network optimizer sections designed in this paper are described in detail as follows. The improved network structure and loss function are described in Sections 3 and 4, respectively.
Data enhancements include random image cropping and image flipping. Image random cropping not only increases the amount of data but also weakens the data noise and increases the stability of the model. Assume that the water droplet regions of the rubber conveyor belt are C1, the nonwater droplet regions are C2, the main features of C1 are C1, F1, G1, and C2 is E2, F2, G2. Assume that background noise is added: C1, C2 randomly add N1, N2, N3, and the image we randomly cropped at this time is as follows:
Image morphology [13] mainly performs closed and open operations on the water droplet binary mask. In order to detect the specific position of the water drop, the difference value between the water drop image and the original image is used to obtain a mask image. The closed operation of the mask map is to first etch and re-expand the image so that the white areas of the water droplet image are connected to each other and the small black holes isolated in the water droplet area are filled to highlight the entire raindrop area. Then, the opening operation of the corrosion after the first expansion is used to eliminate the background bright noise outside the raindrops, selectively retain the water droplets, and obtain the main object water droplets in the image. The image morphology module is placed after the process of obtaining the water droplet mask maps (water droplet binary mask) and image morphology operation can generate sharper water droplet attention map, and the difference between the network learning result and the water droplet mask is generated as small as possible. It is important to generate a water-free image later in the generative network section.
Based on the model (equation (1)), our goal is to obtain the background image B from a given input I. In order to realize the detection and removal of the water droplet image of the rubber conveyor belt, we create the attention map guided by the binary mask M. The threshold is used to determine whether a pixel is part of a water droplet region, and we set the threshold to 50 for all images in our training dataset.
Figure 2 shows the generative network structure, and Figure 3 shows the discriminant network structure, which together constitute the overall architecture of the generative adversarial network proposed in this paper. Input the water drop image of the rubber conveyor belt, and generate a network to generate as realistic a water drop-free image as possible, and generate a water drop attention map for detecting the position of the water drop. The discriminant network will verify that the image generated by the generator network is authentic.
df19127ead