Ling Peng

Aiming at the core challenges of data shortage and high labeling cost in the infrared small target detection task, this paper puts forward an innovative method of integrating generative adversarial network and semi-supervised learning. By designing the small target generative adversarial network (STGAN) based on StyleGAN2, combining the self-attention mechanism and the comparative learning loss, the quality and diversity of generated data are effectively improved. At the same time, the knowledge distillation framework is used to optimize the student model by using the pseudo-tags generated by the teacher model, which significantly improves the detection performance in small sample scenes. The experimental results show that the nIoU (1% labeled data) of 0.709 is achieved by the proposed method on SIRST data set, and the FID index is reduced to 0.45, both reaching the current optimal level. The ablation experiment further verified the key roles of detection loss, self-attention module and REP-PAN feature fusion strategy, and provided a new idea for data expansion and model optimization in the field of small target detection.

Traffic flow counting is an object detection problem. YOLO (" You Only Look Once ") is a popular object detection network. Spiking-YOLO converts the YOLO network from ANN(artificial neural network) to SNN (spiking neural network), which consumes less power than YOLO. However, Spiking-YOLO is difficult to train and has poor accuracy because of overfitting. Dropout and DropConnect are two methods to regularize neural network models and have produced cutting-edge outcomes in many benchmarks. The two drop approaches are first applied to Spiking-YOLO. Considering that a higher membrane potential of a biological neuron implies a higher probability of neural activation, we propose three adaptive drop algorithms—Spiking-YOLO with adaptive Dropout, Spiking-YOLO with adaptive DropConnect, and Spiking-YOLO with group adaptive Drop, which adaptively adjusts the keep probability for training SNNs. The experimental results show that the adaptive algorithms have fast convergence speed with high detection accuracy and mAP value, and work well in the case of high-speed streaming video, making them suitable for online learning scenarios.

Object detection techniques are used to determine object categories and localization in images. Different tasks have different requirements for detection efficiency and accuracy, which leads to high efficiency one-stage object detection methods and high-precision two-stage object detection methods. As the main category of difficult detection objects, small objects have the characteristics of few pixels, wide distribution, and complex background. Many serious tasks in special environments, such as industrial defect detection, large-scale pedestrian statistics, remote sensing image recognition, etc. require the model to have high accuracy for the detection of difficult objects. Aiming at the problem of information loss of small objects in deep network features, this paper proposes a feature levels selection method for backbone network based on effective receptive fields. By moving down the feature level of the backbone network, the size of the effective receptive field corresponding to the selected hierarchical features is matched with the scale distribution of the small objects to be detected to obtain the best small object detection performance. According to the effective receptive field size (ERF), the appropriate Shallow Feature Selection (SFS) is selected to participate in the subsequent detection process to ensure that accurate and sufficient small target information is extracted from the backbone network.Based on one-stage and two-stage object detection methods, this paper proposes improvement measures for small object detection from the data level, network level and training level, and designs experiments and validates them.