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.

Recent advances in deep convolution neural networks (CNNs) boost the development of video salient object detection (SOD), and many remarkable deep-CNNs video SOD models have been proposed. However, many existing deep-CNNs video SOD models still suffer from coarse boundaries of the salient object, which may be attributed to the loss of high-frequency information. The traditional graph-based video SOD models can preserve object boundaries well by conducting superpixels/supervoxels segmentation in advance, but they perform weaker in highlighting the whole object than the latest deep-CNNs models, limited by heuristic graph clustering algorithms. To tackle this problem, we find a new way to address this issue under the framework of graph convolution networks (GCNs), taking advantage of graph model and deep neural network. Specifically, a superpixel-level spatiotemporal graph is first constructed among multiple frame-pairs by exploiting the motion cues implied in the frame-pairs. Then the graph data is imported into the devised multi-stream attention-aware GCN, where a novel Edge-Gated graph convolution (GC) operation is proposed to boost the saliency information aggregation on the graph data. A novel attention module is designed to encode the spatiotemporal sematic information via adaptive selection of graph nodes and fusion of the static-specific and the motion-specific graph embedding. Finally, a smoothness-aware regularization term is proposed to enhance the uniformity of salient object. Graph nodes (superpixels) inherently belonging to the same class will be ideally clustered together in the learned embedding space. Extensive experiments have been conducted on three widely used datasets. Compared with fourteen state-of-the-art video SOD models, our proposed method can well retain the salient object boundaries and possess a strong learning ability, which shows that this work is a good practice for designing GCNs for video SOD.

Convolutional neural networks (CNN)-based salient object detection (SOD) models have achieved promising performance in optical remote sensing images (ORSIs) in recent years. However, the restriction concerning the local sliding window operation of CNN has caused many existing CNN-based ORSI SOD models to still struggle with learning long-range relationships. To this end, a novel transformer framework is proposed for ORSI SOD which is inspired by the powerful global dependency relationships of transformer networks. This is the first attempt to explore global and local details using transformer architecture for SOD in ORSIs. Concretely, we design an adaptive spatial tokenization transformer encoder to extract global-local features, which can accurately sparsify tokens for each input image and achieve competitive performance in ORSI SOD tasks. Then, a specific dense token aggregation decoder (DTAD) is proposed to generate saliency results, including three cascade decoders to integrate the global-local tokens and contextual dependencies. Extensive experiments indicate that the proposed model greatly surpasses twenty state-of-the-art (SOTA) SOD approaches on two standard ORSI SOD datasets under seven evaluation metrics. We also report a comparison results to demonstrate the generalization capacity on the latest challenging ORSI datasets. In addition, we validate the contributions of different modules through a series of ablation analyses, especially the proposed adaptive spatial tokenization module (ASTM), which can halve the computational budget.

We present a novel, robust estimation method to distinguish salient objects from complicated, dynamic backgrounds in videos. In this method, we propose a novel approach to model motion energy based on motion magnitude, motion orientation, gradient flow field, and spatial gradient of the video frame. Furthermore, an effective spatiotemporal objectness map is also proposed to estimate a compact object-like region in the current video frame leveraging both the objectness proposals and the saliency map of the previous frame. Then the current video frame is oversegmented into the granularity of superpixels using the simple linear iterative clustering algorithm. Each superpixel is designated as a node of a graph. The similarity between adjacent superpixels will be assigned as the weight of an edge that connects these two nodes. The feature values of motion energy and spatiotemporal objectness within each superpixel will be averaged respectively, and used to graphically cluster similar superpixels to form the detected salient object. Extensive experiments comparing this proposed new method against twelve existing salient object detection (SOD) methods have been performed using the benchmark datasets unconstrained video saliency detection and densely annotated video segmentation. Superior performance of this proposed SOD method has been observed through three well-known performance metrics: precision-recall curves, F-measure curves, and the mean absolute error.

Multi-agent reinforcement learning (MARL) has become more and more popular over recent decades, and the need for high-level cooperation is increasing every day because of the complexity of the real-world environment. However, the multi-agent credit assignment problem that serves as the main obstacle to high-level coordination is still not addressed properly. Though lots of methods have been proposed, none of them have thought to perform credit assignments across multi-levels. In this paper, we aim to propose an approach to learning a better credit assignment scheme by credit assignment across multi-levels. First, we propose a hierarchical model that consists of the manager level and the worker level. The manager level incorporates the dilated Gated Recurrent Unit (GRU) to focus on high-level plans and the worker level uses GRU to execute primitive actions conditioned on high-level plans. Then, one centralized critic is designed for each level to learn each level’s credit assignment scheme. To this end, we construct a novel hierarchical MARL algorithm, named MLCA, which can achieve multi-level credit assignment. We also conduct experiments on three classical and challenging tasks to demonstrate the performance of the proposed algorithm against three baseline methods. The results show that our method gains great performance improvement across all maps that require high-level cooperation.