Introduction

Infrared temperature measurement is a mature and dynamic technology that is widely used in many industries and organizations. It is useful for measuring temperature in some typical situations, i.e. when the object to be measured is moving or when a fast response is required1. In order to implement a fully automated thermal imaging analysis system for billet temperature, the location of the billet must first be accurately identified in the infrared image of the billet so that the corresponding temperature value can be fitted based on the pixel value. However, infrared imaging suffers from defects such as blurred edges, low contrast and uneven intensity, resulting in a limited segmentation accuracy for the target2.

At present, the specialized research for infrared image segmentation processing is significantly less than that for visible images, and is dominated by traditional segmentation algorithms, e.g., Zhou3 proposed an infrared image segmentation algorithm based on Otsu and genetic algorithm, but the algorithm has multiple thresholding calculations and the processing is relatively complicated. Ochoa4 also used threshold information to segment infrared images and applied it to fault detection of motors, but the segmentation effect for the target was slightly rough. Yin5 proposed a dual-even morphological gradient-based edge detection operator to identify and diagnose inferior insulators, and Wang6 designed a spiking neural network using the properties of spiking neurons to implement edge detection on infrared images of high-voltage equipment. both of the above methods have made certain contributions to edge detection algorithms for infrared images,but the segmentation effect is poor in cases where the gradient between the target and the background is small. Yu et al.7 successfully extracted the target information of IR images based on the region and gradient information of the image, but the algorithm used fuzzy clustering method in extracting the region information, which is more complicated to calculate and difficult to achieve real-time. Most of the above methods use traditional segmentation algorithms in achieving segmentation of infrared targets based on are using low-level features of images (information such as color, texture and shape), which are difficult to apply to billet identification in complex environments in heating furnaces.

In recent years, with the rapid development of computer processing technology, deep learning technology has been widely used in image recognition, semantic segmentation, target detection, and other fields.Different from traditional segmentation methods, the goal of semantic segmentation based on deep learning is to predict the class label of each pixel in the image8, and automatically learn features in various scenarios through a large number of sample training, so it has better generalization capabilities and robustness9.

To meet the needs of the heating furnace control system and field personnel observation, the real-time and accuracy of the infrared image segmentation of the billet in the furnace is an important indicator of the effectiveness of the segmentation method. Existing methods10,11,12 mainly focus on improving performance. However, achieving real-time performance with low latency is the most critical issue for real applications. At present, there are several ways to improve the speed of network segmentation.Some methods speed up network prediction by reducing the resolution of the input image, such as BiseNet13, DFANet14, etc., but it will lose some spatial information, especially edge information. Some other methods prune the number of feature channels to reduce the computational consumption, such as ENet15, SegNet16, etc., and this way will decrease the feature extraction ability of the network. Another common solution is to use fewer downsamples when extracting features, such as ESPNet17, ERFNet18, etc. These networks have an obvious defect that they cannot achieve sufficient receptive fields. In order to solve the dilemma of real-time semantic segmentation, many improved network architectures have been proposed, such as Spatial Pyramid Pooling (SPP)19, Atrous Spatial Pyramid Pooling (ASPP)12, and other structures that increase the receptive field.

In summary, the key to real-time semantic segmentation is how to obtain a larger receptive field and restore spatial information while maintaining a smaller computational cost. Therefore, this paper proposes a segmentation method based on multi-scale feature fusion. First, the information interaction between all levels of feature maps is strengthened through multi-scale feature extraction, and a larger receptive field and spatial information recovery are obtained. Secondly, Dense Upsampling Convolution(DUC) strategy is used to retain more information during decoding, to improve the accuracy. Finally, Finally, this paper uses the lightweight backbone network ResNet-1820 and deep-wise separable convolution to reduce the computational consumption during feature fusion. Experiments have proved that the network proposed in this paper performs well in segmentation accuracy and efficiency, and achieves a detection speed of 80fps and 94.89\(\%\) mIoU on GTX2080Ti.

Methods

In this paper, we propose a real-time segmentation method for infrared images based on a lightweight network, ResNet-18, for multi-scale feature fusion. The overall structure of this network is shown in Fig. 1. The specific steps are as follows.

Figure 1
figure 1

The network structure of the method in this paper.

Full size image

  • Improve the information interaction between the feature maps at all levels of the network through multi-scale feature fusion, and strengthen the multi-scale expression ability of the network;

  • Use Dense Upsampling Convolution(DUC) strategy to reduce information loss, retain more image feature information, and improve segmentation accuracy;

  • Use hole convolution and asymmetric convolution to obtain a larger image receptive field, more fully integrate the context information of the image, and further improve the accuracy of segmentation;

  • Reduce the number of network parameters through lightweight backbone network and depth-wise separable convolution, speed up model inference, and realize real-time segmentation.

Feature extraction module

The feature extraction module of the method in this paper is shown in Fig. 2. Its main structure uses a lightweight network ResNet-1820, which has four different residual blocks, and each residual block contains two 3 \(\times\) 3 convolutions and one skip connection, ResNet-1820 solves the problem of gradient dispersion and gradient explosion in deep convolutional networks by using the residual structure, but it still has some drawbacks such as small receptive field and single feature extraction size.

In order to further reduce the parameters, expand the perceptual field and improve the segmentation accuracy, in this paper, asymmetric convolution21 and dilated convolution22 are applied in Resblock\(\_\)3 and Resblock\(\_\)4. Asymmetric convolution is the decomposition of a standard two-dimensional convolution into two one-dimensional convolutions, i.e., the traditional n\(\times\)n convolution is decomposed into n\(\times\)1 and 1 \(\times\) n convolutions. This approach has two advantages23:increasing the nonlinearity of the network and improving the discriminative power of the network; and reducing the network parameters and computational effort. Dilated convolution can be considered as convolution with holes. The basic principle is to insert a hole (i.e. a pixel with a value of 0) between each pixel of the normal convolution kernel, increasing the field of view of the network without increasing the number of network parameters.

Figure 2
figure 2

The network structure of the feature extraction module.

Full size image

Muti-scale feature fusion

In semantic segmentation, to obtain a larger receptive field and a smaller computational cost, a down-sampling operation is required24. However, this will lose a large amount of spatial information, especially information related to edges. Based on the above analysis, this paper proposes a method of fusing multi-scale feature information to improve the multi-scale expression ability of the network. Specifically, after each residual block of the backbone network, pooling operations of different scales are performed to generate feature maps of different scales. In order to expand the receptive field and extract richer feature information, the pooling operation is used as the step size \(s = 2^j\), and the convolution kernel size is \(k = 2 \times s+1 = 2^ {j+1} +1, j \in [1,2,3]\), j is the pooling level. For the feature map of \(B \in {\mathbb {B}} ^ {C \times H \times W}\),after the pooling operation, the size of the feature map is \(O^{j} \in {\mathbb {B}}^{C \times \frac{H}{2^{j}}} \times \frac{W}{2^j}\), (H, W)is the height and width of the feature map, and C is the number of channels in the feature map. We perform a 3-level pooling operation on the feature map output by Resblock_1, perform 2-level and 1-level pooling operations on Resblock_2 and Resblcok_3 respectively, and finally aggregate the feature maps with the same resolution. Due to a large number of channels after aggregation, the depth-wise separation convolution25 is used for fusion to reduce the amount of calculation for fusion. By merging the feature information extracted from different layers in the backbone network, the interaction between low-level spatial information and high-level semantic information is strengthened, thereby improving the accuracy of the network.

Dense up-sampling convolution

There are three general up-sampling methods for networks under the semantic segmentation task: bilinear interpolation, de-pooling operation, and deconvolution operation. Among them, bilinear interpolation is not learnable, and detailed information will be lost, while the de-pooling operation and deconvolution operation will lose part of the information, which affects the segmentation accuracy. Based on the above situation, Panqu Wang et al.26proposed dense up-sampling convolution (DUC), which uses channel dimensions to make up for the loss in length and width. Specifically, the input image size is set to H\(\times\)W, the feature map after the feature extraction stage is \(h \times w \times c (h=H/d,w=W/d)\), and convolution is applied to this feature map. The resulting output feature map is \(h \times w \times (d^2 \times L)\), where d is the down-sampling multiple and L is the total number of categories of the segmentation task, and finally through pixel shuffle to H\(\times\)W\(\times\)L.

In this paper, dense up-sampling convolution is used in the decoding part of the network structure, and the up-sampling task is allocated to each layer to fuse the information of different layers. The specific structure is shown in Fig. 3. When sampling on each layer, the feature graph is only restored to the size of the previous layer, and the information of each layer is merged with a small computational consumption, which further avoids the loss of information during decoding, and further improves the segmentation accuracy.

Figure 3
figure 3

Dense up-sampling module structure.

Full size image
Table 1 Experiment results of the effectiveness of the network structure.
Full size table
Table 2 Speed and accuracy analysis.
Full size table

Experiment

Data set preparation

In this study, the infrared images of billets at the exit of the heating furnace are used to intercept frames from the video surveillance data, and a total of 5000 images of billets in different states are intercepted, and a \(512 \times 512\) area in the center of the image (the area contains most of the information of the billets) is intercepted. The intercepted images are annotated to obtain the target images, and the training set, validation set and test set are divided according to the ratio of 7:2:1.

Experiment method

In order to verify the effectiveness of the proposed network, we conducted detailed experiments on an experimental platform configured with GTX2080Ti, cuda 10.0, cudnn 7, and pytorch 1.1.0. the model was configured with an Adam optimizer, a batchsize size of 4, an initial learning rate of 1e−4, and 2000 epochs of training rounds. in order to To enhance the robustness of the network, we used data enhancement methods such as random flip, random Gaussian blur, random brightness adjustment, and set random values of [0.5, 2] as the image scale for scaling.

Effectiveness of the network structure

The method in this paper consists of three main parts: a multiscale feature fusion module, a dense upsampling convolution module, asymmetric convolution, and dilated convolution. In order to verify the effectiveness of the network structure proposed in this paper, we conducted detailed comparison experiments. The experimental results are shown in Table 1. from the experimental results, it can be seen that each optimization strategy proposed in this paper has certain superiority, among which Muti-scale Feature Fusion has obvious effect on the prediction of targets and has significant improvement on the prediction of target boundary regions. In this experiment, the basic network is a traditional U-Shape27 structured network.

Comparison of network inference speed and accuracy

To further verify the segmentation performance, some commonly used real-time segmentation networks are selected for comparison experiments in this paper, including ENet15, EspNet17, ErfNet18 and BiseNet13. in addition, to verify whether the selection of backbone network is reasonable in this paper, the backbone network is replaced with MobileNet28 for comparison experiments under the same training parameters. The experimental results are shown in Table 2, among which, BiseNet13 has the fastest prediction rate, but its prediction effect for the target is poor, especially in the boundary area of the target, ENet15 also has poor prediction effect for the boundary area of the target, and ErfNet18 has the closest prediction result with the method of this paper, but the processing efficiency is less than half of the method of this paper. Figures 4 and 5 show some of the results of the comparison experiments.

Figure 4
figure 4

Comparison of image segmentation results by different methods. (a) Ground truth, (b) ours model, (c) U-shape, (d) ERFNet, (e) ENet, (f) EspNet, (g) mobileNet, (h) BiseNet.

Full size image
Figure 5
figure 5

Other comparison.

Full size image

Conclusions

In this paper, we propose a real-time semantic segmentation method, based on the proposed multi-scale feature fusion strategy and dense up-sampling convolution strategy, and apply it to the segmentation of the infrared image of the heating furnace billet. Our network can achieve excellent segmentation accuracy and efficiency, thereby greatly improving the segmentation effect of billets. Finally, by comparing with other existing segmentation methods, the results clearly show that our method has greater advantages in terms of segmentation accuracy and efficiency, meeting the temperature measurement requirements of real-time and segmentation accuracy.