基于空洞卷积和注意力机制的虾苗计数研究OA

To address the issues of small body size,local aggregation,and occlusion affecting counting accuracy in shrimp fry counting,a shrimp fry counting algorithm based on dilated convolution and attention mechanism was proposed,and a corresponding counting system was developed based on this algorithm.The paper used the shallow network of the DSSD(deconvolutional single shot detector)algorithm to generate dilated convolution,there by expanding the receptive field to obtain multi-scale shrimp fry feature information.Then features from the two output feature layers were selected and a new anchor size was designed to accurately detect shrimp fry.To obtain local key information of shrimp fry,a dual attention mechanism by concatenating CBAM(convolutional block attention module)and SeNet(squeeze-and-excitation networks)was introduced.Additionally,the CIoU loss function was employed to solve the occlusion problem.The results show that the proposed model,when trained on a dataset of shrimp fry images with normal and occluded conditions,achieves better recognition and counting performance.Specifically,it improves the accuracy,precision,recall,and F1-score by 4.26%,0.53%,3.87%,and 2.22%,respectively,compared to the original DSSD algorithm,thus meeting the accuracy requirements of practical scenarios.