Radio Frequency Interference Detection Using Efficient Multi-Scale Convolutional Attention UNet

In constructing the backbone network of our model, we draw inspiration from the multi-scale convolutional attention mechanism (MSCA) proposed by Guo et al. (2022) and the efficient channel attention mechanism (ECA) introduced by Wang et al. (2020). The combination of these two mechanisms provides the ability for channel weight adaptation and multi-scale feature processing in RFI detection. Our model is based on a clear and effective encoder-decoder architecture. For the input data, we initially apply a StemConv block (detailed in Table 1), consisting of a dual-layer 3×3 convolution and residual connection (He et al., 2016), for preliminary processing. This design not only prevents gradient vanishing and stabilizes the training process but also extracts low-level features of RFI from the images, providing richer feature information for subsequent layers. In the subsequent encoding and decoding stages, ECA blocks are used to learn the correlations between different channels, adaptively adjusting the weight of each channel to improve the performance and efficiency of the network. Then, the channels with different weights are input into MSCA blocks to extract multi-scale features of RFI. Throughout the encoding and decoding process, we stack a series of these ECA and MSCA blocks to construct a U-shaped network structure (Ronneberger et al., 2015).

In the encoder part, we replaced the previous max pooling down-sampling method with 2×2 convolutional layers, which reduces both spatial and semantic information loss while decreasing the computational burden. In the decoder part, we restored the original image resolution gradually by using transpose convolution to preserve as much image detail as possible. High-resolution feature maps contain rich details and low-level information, but they may have negative impacts on RFI detection. On the other hand, low-resolution feature maps have higher semantic information but lack specific spatial details. Therefore, in the decoding process, we adopted skip connections similar to Unet to integrate the low-level and high-level RFI features (Ronneberger et al., 2015). Through skip connections, we can combine contextual information and multi-scale information, maintain the spatial structure and detail information of the image, compensate for the spatial information loss caused by downsampling, and improve the expressiveness and detection accuracy of the network. Since skip connections double the number of channels, we used 1×1 convolutional layers to re-model and fuse the cross-channel features in order to reduce the number of channels. In the final output layer, we used a 1×1 convolutional layer to reduce the number of channels to 1. Then, we applied the sigmoid activation function to map all results to a probability range of 0 to 1. By setting the threshold to 0.5, we obtained the output segmentation mask. To prevent overfitting, we also introduced the droppath mechanism (Huang et al., 2016) to enhance the model’s generalization ability. In this study, we set the droppath rate to 0.2. As for our model architecture, the use of a U-shaped structure is something that has been done previously; however, those approaches solely employed single-scale small convolutional kernels in their convolutional neural networks. We have replaced these small, single-scale convolutional kernels with larger, multi-scale ones and have integrated attention mechanisms into the network, which is a novelty not found in the methods we compared with. All hyperparameters in this network architecture were carefully tuned and optimized through extensive experimentation.

When observing an image, we typically pay more attention to the parts that interest us. Similarly, attention mechanisms assist neural networks in better utilizing input information. In our network, after the RFI image is processed by convolutional layers, multiple feature maps are generated at different stages. However, not all of these feature maps contribute equally to RFI detection. We should prioritize the feature maps that are more helpful for detecting RFI. Therefore, we employ an efficient Channel Attention mechanism (ECA) to assign different weights to different feature maps, thereby enhancing RFI detection. The schematic diagram of the ECA module is shown in Fig. 3, and its detailed structure can be found in Table 3. During the operation of this module on the input data, global average pooling is applied first to aggregate the RFI features extracted by the convolution. Then, a one-dimensional convolution with a kernel size of k is used to generate weights for each feature map. These weights are then mapped to the range of 0 to 1 using the Sigmoid function, representing the learned channel attention (Wang et al., 2020). Throughout this process, the number of feature maps remains unchanged. The one-dimensional kernel size, k, can be adaptively determined and calculated using Equation  (1),

$\left|t\right|_{\text{odd}}$ represents the closest odd number to $t$ and $C$ represents the number of feature maps, i.e., the number of channels. In this paper, we set the values of $\gamma$ and $b$ to 2 and 1, respectively.

The detection requirement for the RFI model is to be able to identify the presence of RFI at the pixel level, which is essentially a binary classification semantic segmentation problem. According to recent research, an efficient semantic segmentation model needs to have the capability of multi-scale information interaction(Guo et al., 2022). Given the diverse shapes and features of RFI, the complex RFI patterns necessitate the model to possess the ability to handle multi-scale information.

In response to this requirement, the MSCA block is designed to capture the multi-scale features of RFI. Each MSCA block, as shown in Fig. 4(a), consists of Batch Normalization (BN), an attention module for enhancing key information, and a Feed-Forward Neural Network (FFN). The attention module, as depicted in Fig. 4(b), is composed of a 1×1 convolution, Gaussian Error Linear Unit (GELU) (Hendrycks & Gimpel, 2016)activation function, and an MSCA module. The FFN is composed of a 1×1 convolution, 3×3 depth convolution, and GELU activation function. As illustrated in Fig. 4(c), the local details and multi-scale information of RFI are fused using the multi-branch depthwise separable convolution, and the information from multiple feature channels is integrated using a 1×1 convolution. The MSCA can be mathematically represented as follows:

In these two formulas  (2) and  (3), $A_{m}$ represents attention map and $Out$ represents output. $F$ represents input features, DW-Conv represents depth-separable convolution, $Scale_{i}$,$i\in\left\{0,1,2,3\right\}$ represents the $i$th branch in Fig. 4(c), $Scale_{0}$ is the identity connection, and $\otimes$ represents the element-wise Matrix multiplication operation, the detailed structure of MSCA block is shown in the table 4. We use two depthwise separable strip convolutions in each branch to approximately simulate standard depthwise separable convolutions with large convolution kernels. This is mainly based on two considerations. First, strip convolution is more lightweight and can reduce computational complexity. Secondly, this strip structure has been found to be particularly effective in capturing strip-like features that correspond to certain types of RFI(Hou et al., 2020). Therefore, using strip convolution can make the model more accurate in identifying such strip-like RFI.

The detection results of different models on two different datasets are shown in Fig. 5. The left column of the figure displays the Visibility data and their corresponding Ground truth. The second to last columns depict the flagging results of different models, as well as the difference maps between the model outputs and the Ground truth. From the figure, it can be observed that our model achieves the best flagging performance, with the output masks being closest to the Ground truth. This indicates finer edge detection, lower false detection rate for RFI, and fewer false positives. The difference maps are intended to provide a more intuitive visualization of the discrepancies between the output masks of different models and the Ground truth. By calculating the difference between the Ground truth and the output masks of different models and taking the absolute value, the yellow areas in the maps represent the differences between the predicted masks and the Ground truth. It can be clearly observed from these difference maps that our model has the fewest yellow error regions and the smallest differences with the Ground truth. Table 5 presents the scores of four evaluation metrics (precision, recall, F1 score, and IoU) achieved by our EMSCA-UNet model, U-Net model, RFI-Net model, and R-Net model on the test set. To facilitate understanding, the highest score for each metric is highlighted in bold. Specifically, our EMSCA-UNet model achieves scores of 88.08, 83.62, 85.80, and 75.54 for precision, recall, F1 score, and IoU, respectively. The U-Net model and the RFI-Net model obtain similar scores across these four metrics, but the RFI-Net model slightly outperforms the U-Net model in each metric. However, the R-Net model performs poorly, with scores lower than other models in each metric. Overall, our model achieved the highest scores on all four evaluation metrics, averaging about a 5% improvement over the U-Net model. The performance of RFI-Net is comparable to U-Net, with RFI-Net slightly outperforming U-Net, while the performance of R-Net is the worst.

Table 6 displays the specific scores of each model without the use of data augmentation. Upon comparing Tables 5 and 6, a slight decline in most of the model’s performance metrics can be observed when data augmentation is not applied. This observation substantiates the effectiveness of our implemented data augmentation method in enhancing the model’s generalization capabilities. By comparing the structures of R-Net with the other three models, it can be observed that our EMSCA-UNet and RFI-Net models adopt the U-net model’s U-shaped encoder-decoder architecture and employ multiple skip connections. In contrast, R-Net utilizes the classic residual network structure (ResNet) but only employs a single skip connection. Aside from R-Net, the three other models have different numbers of RFI feature maps extracted by the convolutional layers throughout the entire input-to-output process. In the encoder phase, the number of feature maps gradually increases, reaching up to a maximum of 1024. In the decoder phase, the number of feature maps gradually decreases. Different from the other three models, each layer of R-Net only has a fixed number of 12 feature maps. Consequently, through experimental results, it can be concluded that skip connections and a greater number of feature maps are crucial for our RFI detection task. This further corroborates our previous statement in Section  3.1, emphasizing the necessity of integrating skip connections to mitigate network degradation and information loss caused by downsampling, while more feature maps enable our model to learn a wider range of features. By comparing our EMSCA-Unet and RFI-Net models to the U-Net model, it can be observed that RFI-Net augments U-Net by adding an additional 3×3 convolutional layer and introducing Shortcut Connections at each stage of the encoder and decoder. Hence, RFI-Net and U-Net share a striking structural resemblance, explaining the close experimental results between the two. Our model effectively replaces the first of the two 3×3 convolutional layers in U-Net with an ECA module, and the second with an MSCA module, incorporating Shortcut Connections in the MSCA module. Therefore, in a sense, our comparative study between EMSCA-UNet and U-Net can be considered an ablation study. Comparing the first and second rows of Table 7 with the first row of Table 5, it is evident that the performance of individual MSCA blocks or ECA blocks is inferior to when both are used together. Furthermore, from the first and second rows of Table 7 and the second row of Table 5, it can be observed that replacing both convolutional layers of U-Net with MSCA block or substituting the first convolutional layer with ECA block leads to improved performance compared to the original U-Net. These findings indicate the effectiveness of both modules added, and their combined usage yields better results than individual usage. The use of the Efficient Channel Attention (ECA) module allows our model to focus more on feature maps that are pertinent to RFI detection during the training process. Moreover, by utilizing the Multi-Scale Convolution Attention (MSCA) module, which extracts multi-scale features, the model effectively enhances RFI detection compared to solely adopting single-scale 3×3 convolutions. The subpar performance of R-Net may be attributed to the need for transfer learning, as outlined in their original paper where a significant performance improvement was achieved by training on simulated data and fine-tuning on real data. In addition the third row of Table 7 shows the experimental results obtained for our model using the adam optimizer, while the experimental results in Table 5 are all obtained using the adamW optimizer. By comparing the results with the first row of Table 5 it can be seen that the experimental results obtained using adamW and adam are very close, but using adamW leads to a little very small improvement in the model performance. Although this improvement is very small it shows that it is reasonable to use adamW as our optimizer. Despite our model’s superiority in detection results and various metrics, it does not exhibit an advantage in terms of speed, which is an aspect that can be further optimized in our future work.