Research On Marine Small Target Detection Based On UAV Video Analysis In Complex Environment

With the increasing efforts of maritime trade,cargo transportation and seabed energy exploration all over the world,maritime accidents occur frequently.Rapid location of accidents has become the key to emergency rescue.UAV are widely used in maritime rescue projects because of their small size,high speed and low cost.However,the images taken by UAV have small target imaging,and the marine environment is complex and changeable,resulting in poor imaging effect,which is not conducive to subsequent target detection.To solve these problems,this paper based on convolutional neural network,adding preprocessing operation,and introducing the idea of enhanced receptive field and self attention mechanism respectively.This paper proposes two detection algorithms that can quickly locate the location of shipwreck,The specific research contents are as follows:(1)In order to fit the background of maritime rescue project,this paper collects public data sets and real shooting to make maritime target data sets.The dataset contains 2000 pictures,including ships and drowning people.Because it is difficult to collect relevant data sets on the sea in foggy days and the cost of updating data labels is high,this paper makes fogging on the original images based on atmospheric physical model to simulate two weather scenarios of fog and thick fog on the sea.(2)As the marine environment is complex and changeable,the target object is blocked by fog,which affects the subsequent detection efficiency.We haze removal algorithm based on atmospheric scattering model is used to deal with the original image and the defogging effect and time of the three defogging algorithms are compared.Inputting the defogged image into the backbone network can effectively improve the accuracy of target detection.(3)Aiming at the problem that the traditional target detection algorithm is difficult to capture the feature information of small targets,resulting in the difficulty of small target object recognition.This paper introduces the idea of enhanced receptive field,and proposes an improved receptive field module based on RFB module.In order to reduce the amount of model parameters and computational power consumption,the original large convolution of RFB module is removed in this paper.The improved RFB module is composed of three parallel 1×1 convolutions and three 3×3cavity convolutions to ensure that the receptive field will not increase too many parameters at the same time.The RFB-YOLOV4 Tiny algorithm proposed in this paper is about 4 percentage points higher than YOLOV4 Tiny algorithm.(4)Aiming at the problem that traditional target detection algorithms are vulnerable to noise in the process of feature extraction.This paper introduces the idea of self attention and proposes an improved self attention module based on CBAM module.Firstly,in the channel attention dimension,one-dimensional convolution is used to replace the full connection layer to complete the information interaction between cross channels,so that the parameter is reduced to the constant level.Secondly,in spatial attention,the Dilated Convolution with convolution kernel size of 3 is used to replace the7×7 convolution,which can not only increase the receptive field,but also reduce the amount of parameters.The ECAM-Tiny algorithm proposed in this paper has an improvement of about 6percentage points compared with YOLOV4 Tiny algorithm.This paper presents two detection algorithms that can quickly locate the location of shipwreck:RFB-YOLOV4 and ECAM-Tiny.RFB-YOLOV4 is a detection algorithm proposed by YOLOV4 Tiny by introducing enhanced receptive field,and ECAM-Tiny is a detection algorithm proposed by YOLOV4 Tiny combined with self attention mechanism.