Research On Real-time Detection And Segmentation Of River Crabs In Ponds Based On Underwater Machine Vision

River crab farming is the industry with the largest single-species production value of freshwater fisheries in China,and the precise feeding of bait affects the farming cost-effectiveness directly.However,the variation of bait feeding density depends on the actual distribution of crabs at the bottom of the pond,the existing live crab estimation methods largely rely on manual experience,which cannot achieve accurate estimation of crab density in different water quality,growth cycles,and various areas of the pond.To estimate the distribution of free-sensitive live crabs in dynamic turbid pond environments scientifically,achieving precise bait feeding,the problem of real-time detection and segmentation for pond crabs is investigated utilizing underwater machine vision technology.The specific research contents are as follows:First of all,for the problem of low contrast and information loss associated with underwater low-illumination crab images,a single dark-light image enhancement method based on reflectance restoration learning is proposed.The method introduces enhancement learning into reflectance restoration and illumination attenuation compensation.Using a single dark-light image for supervised learning to eliminate artifact information in the dark-light image,the perception ability of crab information is enhanced,which could provide effective feature information for subsequent advanced visual recognition tasks.Second of all,for the difficulty and slowness of detecting live crabs with multi-scale and multi-posture in aquaculture environments,a real-time lightweight multi-scale object detection network based on convolutional neural networks is proposed.The fast multi-scale feature extraction network is first constructed for live crab targets of different scales by depthwise separable convolution,inverted residual bottleneck structure,and feature pyramid.The unified quantized convolutional neural network framework is then utilized to quantify the error correction of the improved detection network,speed up the computation of convolutional layers,and compress the parameters of fully connected layers.The experiments prove that integrating multiple lightweight network methods in the multi-scale backbone network and category/bounding box prediction layer can effectively improve the efficiency of accurate recognition and location of multi-scale live crabs.Ultimately,for the problem of serious missed/misdetection caused by utilizing rectangular bounding boxes to locate non-structural live crabs(extreme cases such as aquatic plant occlusion,crab overlap,etc.),automatic coarse-to-fine joint detection and instance segmentation network is proposed.The IVo VNet-19-DW,which balances accuracy and power consumption,is first designed as the backbone network of the anchor-free fully convolutional one-stage detector.Then the target bounding box is predicted using pixel-by-pixel regression and center-ness strategy to avoid the complex computation associated with anchor boxes.Finally,the novel spatial attentionguided branch is added to enhance the pixel features of interest,thus on the basis of the predicted coarse yet instance-aware rectangular bounding boxes,the detector is extended into a fine onestage instance segmentation network,which can focus on irregular occlusion objects.Experiments show that the extended one-stage instance segmentation network effectively combines the advantages of object detection bounding box-level instance information and semantic segmentation pixel-level segmentation information.It can realize the high precision instance segmentation of underwater non-structural live crabs while having high computational speed.The above dark-light enhancement,multi-scale detection,and instance segmentation methods provide technical support for non-destructive,rapid,and accurate recognition and statistics of live crab biomass and construction of the pond crab density distribution;provide a critical decision-making basis for accurate variable feeding of automatic feeding boats,which is of great significance for further promoting the intelligent development of the aquaculture industry such as river crab.