Most of the edge devices have restricted computational resource, such as ASIC, FPGA or other embedded systems, which cause an efficient problem for neural network model to run in these hardware platform. Model quantization is an effective optimization technique for convolutional layer inference of neural network at the cost of little accuracy loss. However, most of quantized methods only accelerate the computation of convolutional layer, other layers of a model are still inferred by floating-point calculation. FPGA is not an applicable platform for floating-point calculation. In this paper, a completely quantized method is proposed for inference of neural network on FPGA platform. All the calculation of a model inference is performed by quantized value. More quantization leads to more accuracy loss. In order to preserve accuracy, several techniques are used for different functional layer of the neural model. Such as activation layer uses bitwise operation instead of mutilation, concatenate layer use respective parameter for different input layer. To evaluate the effectiveness and efficiency of the proposed method, we implement a quantized light weight detection network, and deploying it on FPGA platform. The experimental result demonstrates that our quantized method is a very low accuracy loss method and is high efficient for neural network inference on FPGA platform. The proposed quantized inference method is highly beneficial for neural model to deploy on low power consumption devices.
The prevalent deep learning approach achieve a great success in many detection task. However, due to the limited features and complicated background, it is still a challenge to apply it to small target detection in infrared image. In this paper, a novel method based on convolutional neural network is proposed to solve the small target detection problem. Firstly, the image feed to neural network is preprocessed in order to enhance the target characteristic by encompassing space and time information. Then the spatial-temporal datum is used to train a custom designed lightweight network dedicated to small target detection. At last, the well trained model is used for inference of infrared video. Furthermore, several tricks are also employed to improve the efficiency of the network so that it is able to operate in real time .The experimental result demonstrate the presented method have achieved decent performance on small target detection task.
A panoramic surveillance system is designed to achieve continuous monitoring of the surrounding environment. The image acquisition module of the system is composed of five fixed-focal-length cameras and one variable-focal-length camera, which realizes 360 degree environmental surveillance. An adaptive threshold is used to dynamically update the background template in order to better accommodate various weather changes. Further, a pixel-level video moving target detection algorithm is applied to effectively detect whether an intruding target exists and determine the direction of the target. It shows the advantages of less computation and preferable detection accuracy. Once an intrusive target is found, the deep convolution neural network SSD is employed to recognize the specific target quickly. As common sense, visual object tracking is one of the most attractive issue in computer vision. Recently, deep neural network has been widely developed in object tracking and shown great achievement. Here, we propose an end-to-end lightweight siamese convolution neural network to achieve fast and robust target tracking. The experiment result shows panoramic surveillance system can effectively and robustly perform security tasks such as panoramic imaging, target recognition and fast target tracking. At the same time, the deep convolution neural network can recognize and track the target accurately and quickly, which meets the real-time and accuracy requirements of practical task.
This paper proposed a fast human action recognition algorithm which utilized two features that can be described as iconic posture and fast moving. At first, a human detection algorithm is used to detect human object in every frame. Then regions marked as human are sent into a trained deep classification network to match trained iconic postures in key frame. Then several frames before key frame and after key frame are examined by frame differences, which are used to compensate background movement and perform human motion speed judgment. After the key frame pinning and speed judgment, the final recognition results are determined.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.