FPGA Accelerated System For Fish Embryo Detection

Target detection of fish embryos plays an important role in fish farming.And FPGA,as a programmable gate array,provides relevant hardware support for the research of fish embryo detection.Deep learning is widely used in various fields because of its ability to extract image features and to fit complex problems.In this paper,we propose(1)a fish embryo target detection model based on the improved YOLOV3 network and(2)a fish embryo detection system based on the improved YOLOV3 network with FPGAs,taking into account the programmability and ease of deployment of FPGAs for the purpose of target detection of fish embryos,combining the computational advantages of deep learning convolutional neural networks.The main contents of this paper are as follows.(1)Embryo culture,as an extremely important part of fishery culture,has high requirements for real-time embryo detection.Although CNN-based target detection algorithms have achieved better results in conventional target detection tasks.In contrast to the detection images of large targets,embryo images are more tiny and fine,containing a lot of embryonic period information in the cell images.Therefore,the existing detection methods spend a lot of time on the task of embryo detection to do some useless calculations and fail to achieve the effect of real-time detection.To solve these problems,the YOLOV3 model is chosen as the original detection model,and an embryo detection model based on the improved YOLOV3 network is proposed in this paper,which consists of three aspects: 1)using VOC2007 data as the pre-training dataset for the YOLOV3 network model for training,and optimizing the trained model;2)using layer pruning and channel pruning(2)the first stage optimization of the model using layer pruning and channel pruning;3)the INT4 optimization of the model based on the TQT quantization strategy.(2)In recent years,significant progress has been made not only in methods but also in theory for target detection in images.However,direct application of these detection methods to production faces new challenges,such as(1)applying deep networks to GPU architectures and deploying them to the field can incur unaffordable costs,and(2)existing CPU+GPU architectures are redundant in terms of their functional modules for deploying neural networks.In this paper,we propose an FPGA-based improved YOLOV3 network for fish embryo detection system using the idea of FPGA acceleration of neural networks and a combination of neural network acceleration on FPGA with Yolo V3 network by borrowing the improved Yolo V3 convolutional neural network in Chapter 3.On the basis of choosing YOLOv3 as the basic model,this paper makes the whole system achieve the effect of real-time detection by combining model optimization and hardware acceleration.In this paper,the model pruning method based on scaling factor is combined with the quantization algorithm of trainable threshold,and the YOLOv3 model is channel pruned with 50% layer pruning and 4-bit quantization,which improves the inference speed of the model by nearly 50% but the m AP of the model is only reduced by 4.23,and it is deployed on a 100 MHz XCZU2CG chip.Real-time fish embryo detection of 34.1 frames is achieved,while having 82.49% m AP value,and the whole system consumes only 2.6w of power during operation.The results show that the FPGA-based real-time fish embryo detection system proposed in this paper can fully meet the needs of today’s aquaculture embedded devices.It is beneficial to promote the transformation of fish embryo analysis from traditional manual methods to AI-based embedded devices and facilitate the rapid analysis of fish embryos.