The Convolutional Neural Network (CNN) has been used in many fields and has achieved remarkable results, such as image classification, face detection, and speech recognition. Compared to GPU (graphics processing unit) and ASIC, a FPGA (field programmable gate array)-based CNN accelerator has great advantages due to its low power consumption and reconfigurable property. However, FPGA’s extremely limited resources and CNN’s huge amount of parameters and computational complexity pose great challenges to the design. Based on the ZYNQ heterogeneous platform and the coordination of resource and bandwidth issues with the roofline model, the CNN accelerator we designed can accelerate both standard convolution and depthwise separable convolution with a high hardware resource rate. The accelerator can handle network layers of different scales through parameter configuration and maximizes bandwidth and achieves full pipelined by using a data stream interface and ping-pong on-chip cache. The experimental results show that the accelerator designed in this paper can achieve 17.11GOPS for 32bit floating point when it can also accelerate depthwise separable convolution, which has obvious advantages compared with other designs.

The breakthrough of deep learning has started a technological revolution in various areas such as object identification, image/video recognition and semantic segmentation. Neural network, which is one of representative applications of deep learning, has been widely used and developed many efficient models. However, the edge implementation of neural network inference is restricted because of conflicts between the high computation and storage complexity and resource-limited hardware platforms in applications scenarios. In this paper, we research neural networks which are involved in the acceleration on FPGA-based platforms. The architecture of networks and characteristics of FPGA are analyzed, compared and summarized, as well as their influence on acceleration tasks. Based on the analysis, we generalize the acceleration strategies into five aspects—computing complexity, computing parallelism, data reuse, pruning and quantization. Then previous works on neural network acceleration are introduced following these topics. We summarize how to design a technical route for practical applications based on these strategies. Challenges in the path are discussed to provide guidance for future work.

Autonomous navigation of Unmanned Aerial Vehicles (UAVs) based on deep reinforcement learning (DRL) has made great progress. However, most studies assume relatively simple task scenarios and do not consider the impact of complex task scenarios on UAV flight performance. This paper proposes a DRL-based autonomous navigation algorithm for UAVs, which enables autonomous path planning for UAVs in high-density and highly dynamic environments. This algorithm proposes a state space representation method that contains position information and angle information by analyzing the impact of UAV position changes and angle changes on navigation performance in complex environments. In addition, a dynamic reward function is constructed based on a non-sparse reward function to balance the agent’s conservative behavior and exploratory behavior during the model training process. The results of multiple comparative experiments show that the proposed algorithm not only has the best autonomous navigation performance but also has the optimal flight efficiency in complex environments.

Computer vision based detection approaches are widely employed to detect or calibrate different types of meters nowadays. However, traditional detection algorithms suffer drawbacks in accuracy and adaptability upon detecting various types of automobile dashboards. Plenty of parameters of these algorithms need to be tuned to suit certain types of dashboards. Besides, theses algorithms cannot automatically read the speed value which requires manual setting operations. In this paper, a novel approach is presented to adaptively detect different types of automobile dashboards. The contour analysis based method is first implemented to extract the connected component of the pointer. A robust character segmentation classifier, which is designed by cascading histogram of oriented gradients (HOG) / support vector machine (SVM) binary classifier, character filter as well as HOG/Multi-class SVM (MSVM) digit classifier, is then proposed to recognize digit characters on the dashboard. Simultaneously, tick marks are then extracted based on recognition results. Finally, Newton interpolation linear relationship is established to diagnose the potential responding errors of the pointer. The experimental results show that the pointer extraction method is robust to interferences caused by connected components of digits and also that the established character segmentation classifier has a more accurate detection result. Furthermore, compared with similar algorithms, it has a significant advantage in detecting a vast majority of different dashboards without manual tuning of the parameters.

Infrared and visible image matching methods have been rising in popularity with the emergence of more kinds of sensors, which provide more applications in visual navigation, precision guidance, image fusion, and medical image analysis. In such applications, image matching is utilized for location, fusion, image analysis, and so on. In this paper, an infrared and visible image matching approach, based on distinct wavelength phase congruency (DWPC) and log-Gabor filters, is proposed. Furthermore, this method is modified for non-linear image matching with different physical wavelengths. Phase congruency (PC) theory is utilized to obtain PC images with intrinsic and affluent image features for images containing complex intensity changes or noise. Then, the maximum and minimum moments of the PC images are computed to obtain the corners in the matched images. In order to obtain the descriptors, log-Gabor filters are utilized and overlapping subregions are extracted in a neighborhood of certain pixels. In order to improve the accuracy of the algorithm, the moments of PCs in the original image and a Gaussian smoothed image are combined to detect the corners. Meanwhile, it is improper that the two matched images have the same PC wavelengths, due to the images having different physical wavelengths. Thus, in the experiment, the wavelength of the PC is changed for different physical wavelengths. For realistic application, BiDimRegression method is proposed to compute the similarity between two points set in infrared and visible images. The proposed approach is evaluated on four data sets with 237 pairs of visible and infrared images, and its performance is compared with state-of-the-art approaches: the edge-oriented histogram descriptor (EHD), phase congruency edge-oriented histogram descriptor (PCEHD), and log-Gabor histogram descriptor (LGHD) algorithms. The experimental results indicate that the accuracy rate of the proposed approach is 50% higher than the traditional approaches in infrared and visible images.