Target detection is one of the most important research directions in computer vision. Recently, a variety of target detection algorithms have been proposed. Since the targets have varying sizes in a scene, it is essential to be able to detect the targets at different scales. To improve the detection performance of targets with different sizes, a multi-scale target detection algorithm was proposed involving improved YOLO (You Only Look Once) V3. The main contributions of our work include: (1) a mathematical derivation method based on Intersection over Union (IOU) was proposed to select the number and the aspect ratio dimensions of the candidate anchor boxes for each scale of the improved YOLO V3; (2) To further improve the detection performance of the network, the detection scales of YOLO V3 have been extended from 3 to 4 and the feature fusion target detection layer downsampled by 4× is established to detect the small targets; (3) To avoid gradient fading and enhance the reuse of the features, the six convolutional layers in front of the output detection layer are transformed into two residual units. The experimental results upon PASCAL VOC dataset and KITTI dataset show that the proposed method has obtained better performance than other state-of-the-art target detection algorithms.

Target detection based on deep learning are developing rapidly. However, small target detection is still a challenge. In this paper, a simple and efficient network for small target detection is proposed. We put forward to improve the detection performance of the small targets in three aspects. Firstly, as the contextual information is important to detect the small targets, we proposed to use ‘dilated module’ to expand the receptive field without loss of resolution or coverage. Secondly, we applied feature fusion in different dilated modules to improve the ability of network in detecting small targets. Finally, we used ‘passthrough module’ to get the finer-grained information from the earlier layer and combined it with the semantic information from the deeper layer. To improve the detection speed of the network, it is proposed to use 1×1 convolution to reduce the dimension of the network. We composed small vehicle dataset based on VEDAI dataset and DOTA dataset respectively and also analyzed the distribution of the small targets in each dataset. To evaluate the performance of the proposed network, we trained the model on the dataset above and compared with the state-of-the-art target detection algorithms, our approach achieved 80.16% average precision (AP) on VEDAI dataset and 88.63% AP on DOTA dataset and the frames per second (FPS) is 75.4. The AP of our network is much better than the result of tiny YOLO V3 and is nearly the same as the result of YOLO V3. However, the FPS of our network is almost the same as that of tiny YOLO V3.

Deep learning-based Synthetic Aperture Radar (SAR) ship detection methods are significant in signal processing and radar imaging. However, these approaches always require large-scale SAR ship images with labels to train the model. Due to the inaccessibility of SAR sensors, it is difficult to acquire enough SAR images. Annotating ship targets also demands resources and manpower. To tackle this issue, we propose a novel SAR image generation method named SARGAN for SAR ship detection task. Given the position and category, SARGAN can generate realistic SAR images with SAR ship targets, land and background in the desired location. In the SARGAN, there are five components: target encoder, scene constructor, SAR image generator, and target and image discriminators. The target encoder is introduced to predict the latent vector for each target while the scene constructor integrates all targets in the entire scene using convolutional LSTM. We improve the structure of the SAR image generator by adding operations to generate high-quality images. The image and target discriminators are responsible for distinguishing between real and fake samples, with the latter also predicting the category. To promote the generation of diverse and realistic SAR ship images, multiple loss functions are employed for training. Additionally, we have annotated the lands and background in the HRSID dataset and combined them with labeled ships to create a new dataset for training and testing of SARGAN. Extensive experiments demonstrate that SARGAN outperforms other SAR image generation methods, and the generated SAR ship images are highly conducive for SAR ship detection task.

In the task of Synthetic Aperture Radar (SAR) ship detection, ship targets exhibit arbitrary orientations and are closely aligned. In recent years, oriented bounding box (OBB) based detectors have gained attention as a solution to the significant overlap issue present in horizontal bounding box (HBB) based detectors. Due to the boundary discontinuity issue with current mainstream OBB methods, detectors utilizing the polar coordinate system to define OBB have been introduced. However, currently available methods are usually complicated to encode and decode, and do not take into account the complex background of SAR images. This results in the complexity of neural networks and inaccurate predictions. In this paper, we propose a novel five-parameter polar coordinate dense regression detector (FPDDet) that only uses a centroid, a mapped polar diameter, and two polar angles to deal with the overlap problem of HBB and the boundary discontinuity problem of OBB. Meanwhile, we introduce a dense regression strategy based on covariance-adaptive rotated Gaussian heatmap to dynamically assign ship target samples in response to large-scale variation of ship targets in SAR images, and we suggest a dense regression heatmap loss function to better match our dense regression strategy. In addition, we design a feature enhancement module to enhance the target features while weakening the background interference, aiming to cope with the severe noise pollution of SAR images. Experimental results show that our FPDDet achieves the state-of-the-art performance on Rotating SAR Ship Detection Dataset (RSSDD) and Rotated Ship Detection Dataset (RSDD). Compared to previous best results, mAP has been improved by 1.2% and 1.71%, respectively.