Research On Aerial Image Target Detection Technology Based On MPSoC

With the rapid development of aerial optical imaging technology in recent years,airborne image processing system requires more and more high precision and speed of target detection.Traditional target detection algorithms have been unable to meet the requirements.At the same time,the target detection algorithm based on deep learning has been widely concerned by the academic community due to its better performance.However,such algorithms often have more parameters,higher time complexity and are difficult to be transplanted in mobile terminal.Aiming at the above problems,this paper studies the aerial image target detection technology based on MPSoC platform.The main research work is as follows:(1)The present situation of target detection algorithms at home and abroad is investigated.Firstly,the working flow and working principle of the traditional target detection algorithm are investigated,and the reason why the detection accuracy and detection speed are difficult to be improved is analyzed.Then,the characteristics and development context of the two-stage network represented by R-CNN and the single-stage network represented by YOLO in the deep learning-based target detection algorithm are investigated.(2)YOLOV3,the basic network selected in this paper,is analyzed in depth.From the basic building units to the entire residual network construction.Various techniques existing in YOLOV3 network are analyzed,including the significance of Anchor Box for reducing computation and the role of FPN for multi-scale detection.Finally,the loss function is analyzed and the positive effect of each part on target detection is clarified.(3)The image before detection is enhanced.The types and causes of noise in the process of image acquisition and transmission are analyzed.The denoising effect of several denoising methods is verified experimentally and compared from two aspects of objective evaluation index and subjective evaluation index.Finally,the method of median filter and Gaussian low pass filter is determined.At the same time,the causes of motion blur are analyzed.The Wiener filter,Lucy-Richardision algorithm and a series of fuzzy removal algorithms are compared.Finally,Lucy-Richardision algorithm is selected to remove motion blur.(4)The YOLOV3 model was optimized.Firstly,in the case of aerial image target detection,the traditional Euclidean distance is replaced by the average IOU,and the improved K-means clustering algorithm is used to re-cluster the Anchor Box.After that,it deals with the problem that the feature graph in YOLOV3 network is small and the feature of small target cannot be well expressed,thus affecting the detection effect.The step size of the convolution kernel is adjusted so that the size of the adjusted feature graph is doubled.The experiment showed that the optimized model improved the MAP of the first 10 categories of the VISDRONE data set by1.3%,greatly reduced the false detection rate and improved the recall rate.(5)The algorithm was transplanted by MOSo C.The environment of the hardware platform was built,including the compilation and installation of Open CV,the development of V4L2 interface and the optimization of the main program.At the same time,the YOLOV3 algorithm was further optimized.The influence of each convolutional layer on the network performance was analyzed experimentally,and the model was pruned according to its sensitivity.The experimental results show that the speed of model detection is doubled after pruning,and the volume is reduced to 37%.It basically meets the design requirements of aerial image target detection,and provides a feasible scheme for the implementation of deep learning algorithm in MPSoC platform.