Research On Key Technologies Of Underwater Target Recognition Based On Zynq Platform

With the development of underwater exploration,the demand for underwater application is expanding.In the military field,the detection of potential dangerous targets such as mine in the water environment is an important link to ensure underwater safety.Since the traditional methods such as template matching and SVM have the problems of low recognition accuracy and complex reconstruction process in underwater target recognition.It is necessary to find better methods to improve the recognition rate and the adaptability to image transformation.With the arrival of the era of big data,deep learning method has replaced the traditional target recognition method gradually.As an important branch of deep learning algorithm,convolutional neural network(CNN)has been gradually applied to underwater target recognition.CNN has a good adaptability to image position translation,angle tilt,spatial rotation and other forms of transformation.The network adopts end-to-end mode to avoid complex reconstruction process.However,in order to obtain higher accuracy,the amount of calculation and parameters of CNN models increase rapidly.In the process of underwater detection,a great deal of calculation and parameter storage will pose a great challenge to the power and real-time performance of the equipment.In order to solve the above problems,this paper select MobileNet as the basic framework of this research and propose a parallel acceleration framework which is based on depthwise separable convolution.This parallel acceleration framework can be reused and reconfigured.In this way,the resource and the power consumption can be reduced.This topic uses ZedBoard as the hardware implementation platform.There are three key technologies and difficulties: Firstly,this topic designs a special parallel acceleration unit based on deep separable convolution,which is distinguish from the standard convolution-based parallel computing methods in most studies.On this basis,the whole operation of the architecture is realized in the form of reuse.Secondly,through configurable design,a depthwise separable convolution layer can support multi-layer operations with different configurations.Finally,this topic optimizes the timing of the architecture and realizes the maximum parallel operation between convolution layers.In this way,the real-time performance of the system and the resource utilization of hardware acceleration unit can be improve.In this paper,we use the parallel computing method to replace the serial computing method,which increases speed by 2.83 times.Basing on the uncompressed arthtecture,we compressed the architecture by 0.25 and 0.5 times.Under the compression ratio of0.5 and 0.25,the system frame rate can reach 1.38 fps and 5.52 fps.Compared with other CNN architecture implemented in Zynq7z045,our design uses less resource and has lower power consumption.As the degree of architecture compression increases,our design also has advantages in computing speed.This topic has certain application prospects in underwater target tracking and underwater mine clearance.