Research On Image Enhancement And Real-time Detection Methods For Pond Cultured Crab

River crabs are an extremely important type of aquaculture animal in China’s fish farming species.River crab meat is tender and tasty,and is loved by a wide range of diners.The total value of river crab farming in China is currently around 55 billion,making it the largest industry in freshwater fisheries in terms of production value and high economic value.At present,river crab farming still relies on fishermen’s long-term feeding experience to determine the amount of bait to be fed,which is overly manual and does not accurately determine the amount of bait to be fed and the area based on the growth period and distribution of river crabs.Therefore,in order to meet the need of identifying the number of river crabs required for accurate bait feeding in pond culture environment,this paper uses underwater image processing algorithms and deep learning target detection techniques to investigate the problem of target detection of river crabs in ponds:An image enhancement method based on multi-scale pyramid fusion is proposed for underwater river crab images with blurred and uneven brightness characteristics.The method first applies the dark channel a priori based on blue-green channel and CLAHE to pre-process the original underwater image for defogging and contrast enhancement,then calculates the contrast weight,saliency weight and saturation weight for the two pre-processed images,followed by normalization,and inputs the normalized weight map into Gaussian pyramid decomposition,and inputs the defogged image and contrast enhanced image into Laplace After fusion and reconstruction,the underwater river crab image enhancement is completed.Finally,the UIQM and UCIQE metrics demonstrate the effectiveness of the algorithm in enhancing the image feature details,which can provide effective information for the subsequent target detection process.In order to achieve fast and accurate recognition of underwater multi-scale and occluded river crab images,this paper proposes a target detection network Mobile CenterNet for pond farming river crabs,which uses the anchorless one-stage network CenterNet as the base model to avoid tedious post-processing and improve detection speed.A MobileNetV2 model with a Coordinate Attention mechanism is used as the backbone network to extract river crab features,which not only achieves light weight but also improves the focus of the model on river crabrelated features.A feature fusion module was designed to extract multi-scale feature map information,and Atrous Spatial Pyramid Pooling(ASPP)was added to fuse the contextual information of different sensory fields.The experimental results show that Mobile CenterNet detection accuracy AP and F1 values reach 97.86% and 97.94%,the model size is only 24.46 M,the detection speed reaches 48.18 frames/s,and the storage memory required for model training is reduced by 81% compared with the baseline model ResNet18-CenterNet,and the AP under the river crab dataset increases by 3.2%.The test results show that the proposed method can achieve fast and accurate detection of river crabs in ponds,providing an effective basis for real time monitoring and scientific baiting in river crab farming.So as to be applied on the actual bait ship embedded computing platform,an underwater river crab target detection system based on the NVIDIA Jetson Nano embedded hardware platform was built.The porting of target detection algorithm and hardware acceleration optimization were completed.Based on PyQt5,the visual interactive interface is integrated with image enhancement and target detection functions to identify and locate river crab targets and display coordinate information in real time.The practicality and rationality of the system is verified through experiments.