Research On Improved Light SOLOv2 Fish Image Instance Segmentation Method

Intelligent farming has become the core driving force of the development of China’s fish farming industry,and this intelligent farming mode will gradually replace the previous traditional farming mode,and the aquaculture industry enters the era of wisdom.Fish farming density is a key factor in the culture process,too small will affect fish output value and economic benefits,too large will affect water quality and fish growth performance.Fish instance segmentation is an important prerequisite for the realization of refined culture,which can provide a basis for the assessment of fish body length,body mass and growth state.However,at the present stage,the instance segmentation network of farmed fish images is difficult to accurately segment the fish bodies of different scales in the figure due to the complex underwater environment and uneven farming density,leading to the problems of low image segmentation accuracy and poor extraction and fusion of small target fish feature information and edge information,and there is still some room for improvement in the segmentation accuracy and network structure.In order to solve the above-mentioned problems of underwater images,a series of research works on accurate fish body cutting based on deep learning are carried out in this paper,which are mainly as follows:(1)An Instance Segmentation Model Based on Deformable Convolutional(DCN)Feature Extraction.Aiming at the problem that the image of Takifugu rubripeshas a large interference of similar underwater background information in the culture environment,resulting in low segmentation accuracy of fish image,an instance segmentation model D-SOLOv2 with deformable convolution is proposed.In order to improve the feature extraction ability and segmentation accuracy of the model,deformable convolution is integrated in the backbone network to expand the receptive field of the model,so as to change the geometry of different fish shapes such as shape and size to be close to each other,and improve the network segmentation ability.In order to verify the effectiveness of the improvement strategy proposed in this paper,a comparative experiment was designed,and the exclusive dataset of farmed Takifugu rubripes was used in the experiment,and in order to expand the dataset sample,the dataset was enriched by data enhancement to prevent model overfitting.The experimental results show that the average segmentation accuracy of the improved SOLOv2 model is 64.4%,which is 2.9% higher than that of the original SOLOv2 model,verifying the effectiveness of deformable convolution in improving the segmentation accuracy of Takifugu rubripes.(2)A small target instance segmentation model based on SimAM attention mechanism.Aiming at the problems of poor image segmentation,false detection and missed detection of small target images due to high underwater aquaculture density,overlapping fish occlusion,and difficulty in extracting small target fish body features,a DS-SOLOv2 instance segmentation method based on SimAM attention mechanism and deformable convolution is proposed.In order to improve the small target image feature extraction ability,the parameter-free attention mechanism SimAM is introduced,which allows the model to focus on the key information of Takifugu rubripes without increasing the parameters and complexity of the network model,improves the detail perception at the boundary,and enhances the model’s recognition accuracy of small targets of Takifugu rubripes.The experimental results show that the average accuracy of the improved DS-SOLOv2 model is 65.2%,which is improved by 3.7%,and the segmentation accuracy of small target fish is 46.8%,which verifies the effectiveness of the improved algorithm proposed in this paper.(3)An optimized deployment method based on the Tensor RT model.In order to make the improved DS-SOLOv2 model better applied in real life,the improved DS-SOLOv2 model is deployed on edge devices,but due to the large network parameters of the model,the inference speed during edge device deployment is slow.Therefore,this paper performs network layer vertical fusion and FP16 low-precision quantization of the improved DS-SOLOv2 model based on the Tensor RT framework to optimize the network structure and accelerate it to improve performance and reduce memory requirements.The experimental results show that the improved DS-SOLOv2 model is optimized by the Tensor RT framework,and the inference speed is improved by 2.85 times without affecting the segmentation accuracy.The improved Tensor RT model was deployed to Jetson Nano for testing and analysis,and the inference speed of the optimized model was improved by 0.91 times,thus verifying that the optimized deployment method based on the Tensor RT model proposed in this paper has the advantages of small number of network parameters and high real-time performance,and can cope with the task of real-time fish detection in the factory farming process..