Research On Object Recognition Of Underwater Creature Based On Mask R-CNN

Accurate underwater target recognition is the guarantee for efficient running of underwater vehicles.However,in complex and varying underwater scenes,optical imaging generally has a series of problems such as uneven lighting,low contrast,blue or green tone,and blurred details.In addition,underwater shooting needs a large amount of manpower and money,thus the collection of underwater images is difficult.Hence,high-quality underwater biological datasets are extremely scarce,and the biological categories are not rich enough,which bring huge challenges to underwater object recognition based on optical imaging.In order to deal with the scarcity of high-quality underwater images and the poor quality of underwater imaging which bring challenges to object recognition,this thesis takes the object recognition of underwater creatures as an example,and proposes a method combing image enhancement and Mask R-CNN framework to realize instance segmentation and object recognition of underwater creatures on a small sample dataset.The main research works completed in this article are as follows:(1)Building an underwater creature dataset for instance segmentation and object detection with image augmentation.Aiming at the lack of high-quality underwater creature images and the corresponding problem of over-fitting of deep model: First,an initial dataset consisting of 84 pictures and 501 creatures is established by manual labeling;Then,use image augmentation methods including GAN(generative adversarial networks)to enlarge the dataset to 430 images,including 2262 creatures,which creates conditions for transferring training and freezing training in model training,thus overcome the overfitting problem in the case of small sample datasets.(2)In order to solve the adverse effects of poor image quality on target recognition,an improved MSRCR(multi-scale Retinex with color restoration)image enhancement method is proposed from the perspective of improving target recognition,which greatly improves the perception of underwater images and equalizes the color distribution of underwater images to improve the Mask R-CNN on recognizing underwater creatures.And compare this method with the other image enhancement methods such as dark channel prior,MSRCR,contrast limited adaptive histogram equalization(CLAHE),etc..From the subjective visual perception to traditional evaluation indicators(information entropy,clarity,contrast)and recognition rate improvement of Mask R-CNN,the effectiveness and advancement of the proposed image enhancement algorithm are verified.(3)The multi-target and multi-class target recognition based on Mask R-CNN.First,fine-tuning the standard Mask R-CNN framework to adapt to the lack of big targets in the used dataset,and realizing multi-target multi-class underwater creature target detection and instance segmentation based on image enhancement.The precision is 98.20%,recall is 95.34% and m AP(mean average precision)is 95.09%.Then,this method is compared with YOLOv3(You only look once),SSD(Single Shot Detector)and a SIFT-based(Scale-invariant feature transform)target detection model to further verify the effectiveness.By comparing the recognition results with and without image enhancement,and comparing the proposed target recognition model based on Mask R-CNN and the other models,the effectiveness and superiority of the proposed method are proved.Meanwhile,it is demonstrated that on a small sample dataset and in an underwater scene,appropriate image enhancement algorithms can improve the capacity of Mask R-CNN,and this improvement is proportional to the results of image quality assessments.This thesis provides an important reference for the development of optical vision systems of underwater vehicles,and has broad application prospects in marine biological source surveys,underwater fishing and marine fortification construction.