Synthetic aperture radar (SAR) image target detection methods based on semi-supervised learning, such as the mean teacher framework, have shown promise in diminishing the issue of limited labeled data. However, several challenges exist in current methods. Firstly, data augmentation techniques designed for optical images may not be suitable for SAR images due to differences in imaging methods. Additionally, the contribution of pseudo labels remains constant during the initial retraining stage can lead to degradation in prediction results. Moreover, the low quality of predicted bounding boxes poses a challenge for effective retraining. To address these challenges, we propose an end-to-end semi-supervised detection method based on the mean teacher framework. To enhance the robustness of training, we firstly introduce SAR-specific data augmentation techniques, including multiplicative noise, which effectively increase the diversity of training samples. Secondly, we propose a method that weights the losses of pseudo-labeled data using a hard-sigmoid function, gradually improving the importance of pseudo-labeled data during retraining, thereby alleviating their potential negative impact on the training process. Finally, we propose an IoU-aware subnetwork to incorporate high-quality pseudo-labeled bounding boxes into retraining, allowing them to contribute to network adjustments while mitigating the impact of low-quality samples. Experimental evaluations on publicly available SAR image datasets demonstrate the effectiveness of our proposed method in improving the target detection capability of semi-supervised SAR target detection.

Nowadays, object detection methods based on deep learning are applied more and more to the interpretation of optical remote sensing images. However, the complex background and the wide range of object sizes in remote sensing images increase the difficulty of object detection. In this paper, we improve the detection performance by combining the attention information, and generate adaptive anchor boxes based on the attention map. Specifically, the attention mechanism is introduced into the proposed method to enhance the features of the object regions while reducing the influence of the background. The generated attention map is then used to obtain diverse and adaptable anchor boxes using the guided anchoring method. The generated anchor boxes can match better with the scene and the objects, compared with the traditional proposal boxes. Finally, the modulated feature adaptation module is applied to transform the feature maps to adapt to the diverse anchor boxes. Comprehensive evaluations on the DIOR dataset demonstrate the superiority of the proposed method over the state-of-the-art methods, such as RetinaNet, FCOS and CornerNet. The mean average precision of the proposed method is 4.5% higher than the feature pyramid network. In addition, the ablation experiments are also implemented to further analyze the respective influence of different blocks on the performance improvement.

The recognition of partially occluded targets is a difficult problem in the field of synthetic aperture radar (SAR) target recognition. To eliminate the effect of occlusion, the intuitive idea is to determine the exact location and the size of the occluded area. However, this is very difficult, even impossible in practice. In order to avoid this difficulty and to improve the recognition performance for the partially occluded target, a fusion strategy of the sparse representation (SR) model based on randomly erased images is proposed to recognize the partially occluded target. The proposed method randomly erases some areas many times in both the test samples and the training samples. The erased training samples in each erasure are used to sparsely represent the corresponding erased test sample. Finally, all the SR results are fused to recognize the test sample. The proposed method utilizes random erasure to eliminate the possible occluded region. In addition, this method uses the fusion strategy to overcome under-erasing of the occluded region and erroneous erasure of the unoccluded region. The key parameter of the proposed method is the erasure ratio only. Although the erasure is random, the recognition performance of the method is relatively stable. Therefore, the method can eliminate the influence of occlusion without determining the details of occlusion. The experimental results show that the proposed method is significantly better than the state-of-the-art methods in the case of occlusion. Additionally, the recognition performance of the proposed method is similar to some comparison methods in the case of no occlusion.

Due to the ability of powerful feature representation, deep-learning-based object detection has attracted considerable research attention, and many methods have been proposed for remote sensing images. However, there are still some problems that need to be addressed. In this paper, a novel and effective detection framework based on faster region-based convolutional neural network is designed. Specifically, first, in order to locate the boundaries of large objects and find the missing small objects, DetNet is incorporated into the detection framework as the backbone network. DetNet fixes the spatial resolution in deep layers and adopts dilated bottleneck with convolution projection to increase the divergence between input and output feature maps. Then, the proposed framework uses the backbone network to extract the scene features and region features simultaneously, which are both mapped to feature vectors and then fused together. The feature fusion operation can improve the feature representation of the generated region. Last, to improve the performance of localization, the cascade structure is adopted in the framework. The cascade structure has multiple phases and every phase has independent classifier and regressor. The results obtained from the previous phase are used as the regions of interest in the next phase. Therefore, the multiphase detector can increase the detection accuracy phase by phase. Comprehensive evaluations on a public ten-class object detection dataset demonstrate the effectiveness of the proposed framework. Moreover, ablation experiments are also implemented to show the respective influence of different parts of the framework on the performance improvement.

Partially occluded target recognition is a problem that is rarely studied but must be faced in the field of synthetic aperture radar target recognition. The difficulty of this problem is finding out the location of the occluded area and eliminating the effects of occlusion. In this paper, an adaptive weighting method based on subimage sparse model is proposed to recognize the partially occluded targets. This method divides the test image (query) and the dictionary into subimages and subdictionaries by the same gridding method. Then, the sparse reconstruction error of each subimage is calculated to design weighting coefficients for the subimage. By multiplying the subimages and subdictionaries by the corresponding weighting coefficients, the weighted query and the weighted dictionary are obtained. Finally, the sparse representation recognition method is utilized to classify the weighted query. The core idea of the proposed method is that the sparse reconstruction error of the image is defined as the sum of the sparse reconstruction error of all the image pixels, and the difference of the sparse reconstruction error of the subimage between unoccluded condition and occluded condition is obvious. So the sparse reconstruction error of the subimage can be easily extracted and used as an indicator for determining the occluded area. By setting small weighting coefficients in the occluded area, the influence of occlusion is reduced and the recognition performance of partially occluded target is improved. Experimental results demonstrate that the proposed method leads to better recognition result in comparison to some state-of-the-art methods under the occlusion condition.

In recent years, deep-learning-based methods for remote sensing image interpretation have undergone rapid development, due to the increasing amount of image data and the advanced techniques of machine learning. The abundant spatial and contextual information within the images is helpful to improve the interpretation performance. However, the contextual information is ignored by most of the current deep-learning-based methods. In this letter, we explore the contextual information by taking advantage of the object-to-object relationship. Then, the feature representation of the individual objects can be enhanced. To be specific, we first build a knowledge database which reveals the relationship between different categories and generate a region-to-region graph that indicates the relationship between different regions of interest (RoIs). For each RoI, the features of its related regions are then combined with the original region features, and the fused features are finally used for object detection. The experiments conducted on a public ten-class object detection dataset demonstrate the validity of the proposed method.