Research On Small Object Detection And Recognition Of Underwater Vehicle Under Sample Deficiency

With the development of underwater vehicle technology,the underwater vehicle autonomous operation technology has received more and more attention.The underwater vehicle autonomous operation technology is composed of a number of key technologies,such as underwater environment sensing,visual servo control,path planning,underwater manipulator,and so on.The underwater environment sensing technology is the eye of the underwater vehicle and plays a vital role in the autonomous operation of the underwater vehicle.Underwater object detection and recognition is a key technology of underwater environment sensing.Traditional object detection and recognition algorithms have the disadvantages of low accuracy and poor robustness.The emergence of the Convolutional Neural Network(CNN)has led to the tremendous development in the field of deep learning.Deep learning based object detection and recognition algorithm demonstrated strong capabilities and shows great advantages compared with the traditional method.Therefore,deep learning based underwater object detection and recognition will be the future research trend.The deep learning based object detection and recognition method relies on a large number of training samples.However,it is difficult to obtain a large number of underwater images and labeling them.Therefore,how to train a generalized underwater object detection and recognition model with the insufficient sample?The deep learning based object detection and recognition algorithm has low detection accuracy for small objects and deformable objects.Therefore,how to improve the detection accuracy of the underwater small object and deformable object?To solve the above-mentioned problems,this paper studied the deep learning based underwater object detection and recognition method.The main research contents of this paper can be summarized as the following three parts.1)For the problem of insufficient underwater training samples,three data augmentation methods are proposed.Firstly,marine turbulence has a fuzzy effect on underwater pictures.We using Wiener filtering to reconstruct the image,then use the inverse process of Wiener filtering to add different degrees of marine turbulence to the restored image.Secondly,Underwater images taken at different shooting angles are different.Therefore,we consider using perspective transform to convert the shooting angle of the image.Thirdly,artificial light will cause uneven illumination of underwater images.Therefore,we collected some classic underwater uneven illumination images,then we extract illumination templates of underwater uneven illumination images,finally we use the illumination templates to fuse with the original image.The experiment results prove that through the three data augmentation methods,under the condition of insufficient training data,an underwater object detection model with strong generalization ability is trained.2)Due to the fixed geometry of the CNN,the deep learning based object detection method has the disadvantage of recognizing a deformable object.So,a deformable network is introduced,the deformable convolution module and the deformable ROI pooling module are used to improve the model,improved the deformable characteristics of the model.The experimental results show that the model improves the detection and recognition performance of deformable objects.3)The deep learning based object detection and recognition algorithm uses the high-level feature map for detection and recognition.The high-level feature has low resolution,so the pre-selection box obtains less information.Therefore,it is difficult to classify a small object and regress its bounding box.We consider fusing the feature maps of different layers,pooling the low-level feature map to reduces resolution,and deconvolution processing the high-level feature to improves the resolution,and then merges the low,medium and high-level feature maps.At the same time,we add a set of small pre-selection boxes.The experimental results show that the improved model improves the detection and recognition accuracy of small objects.