Research And Implementation Of Ship Target Detection In The Gaofen Satellite Imagery Based On FPGA

Ship target detection has been the primary focus of computer vision research efforts,which plays an important role both in the military and civilian.Along with the rapid development of earth observation,the high-resolution remote sensing images have been substantially improved.Consequently,we can obtain a lot of resolution of remote sensing images,but massive data of the high-resolution remote sensing images takes place much of the space and has more details of the objection,which lead to the fact the traditional visual functions have been unable to meet the requirements of data processing and analysis.In recent years,the vigorous development of deep learning models has greatly improved the accuracy in object recognition and location detection.However,due to the huge demand of the convolutional neural network for computing capacity and storage of the hardware platform,it is a challenge that building the convolutional neural network applications when deployed under the traditional platform.Although the GPU platform can provide enough computing capacity,we also consider the power consumption in a specific situation.Meanwhile,the FPGA can have more processing speed than CPU and allow the design energy efficiency.Based on this background,this thesis investigates that we design a ship target detection system for the process of Gaofen satellite images objection detection and the results are as follows.(1)Referencing the DenseNet,our object detection model adjusts the backbone based on the YOLOv3,which reduces the computing capacity,and meanwhile,the loss function was modified by replacing the IoU to GIoU to enhance the model performance.Then we analyze the results.(2)The scheme of the ship target detection system is implemented by Node.js and includes the system structure and inner model.The labeling function is implemented by Konva.js that is a Canvas framework that extends the context.(3)Using image processing and mathematical morphology to improve the efficiency of filtrating unqualified images and using the ship target detection system to arrange the task to improve the efficiency of labeling images.(4)Based on the FPGA,this thesis builds an object detection acceleration subsystem platform and introduces the process of transplanting,the scheme of the system,order of call function,thread allocation,and communication interface.Then show the result,examination conclusion,and resource utilization.