【摘要】 提出了一个改进Mask RCNN目标检测算法用以对养殖梭子蟹进行视觉特征测量.通过在养殖区域采集梭子蟹图像,用上位机识别梭子蟹旋转角度以及甲长和甲宽方向,对输出的Mask进行模板修补,提高模板内区域的置信度.通过图像-实景对应关系换算梭子蟹的真实尺寸,并估算其投影面积、甲宽与甲长,结果准确率高于85%.同时,对视觉算法得到的梭子蟹尺寸特征与其体质量进行拟合,引入k-means聚类,实现双模型支持向量回归机(SVR)预测结构.通过差分进化算法对SVR适应度函数进行寻优,设计了随迭代次数、寻优效果同步变化的缩放因子,以及适者更易生存策略的交叉概率因子,以验证改进算法的寻优能力.测试时,对新传入的数据首先进行归一化处理,然后判断所归属的聚类中心,再传至相应的SVR模型进行预测.测试结果相对误差小于18%.更多还原

【Abstract】 The data of farmed crabs mainly rely on manual measurement, which is both labor intensive and resource wasting. In this paper, an improved Mask RCNN size measurement algorithm is proposed, and the original Mask based on probability output from Mask RCNN is modified according to the template obtained in advance to achieve more accurate size estimation. First, crab images are captured by using cameras, which are then segmented with the Mask RCNN algorithm. After applying a second-order moment algorithm on the mask, the template is transferred and rotated into the center of the original mask, and the corresponding part is modified by adding and subtracted benefits, so as to obtain a more accurate image of the carapace. The real size of the crab is converted through the image-real scene mapping, and the area, width and length of the crab carapace are estimated with an accuracy of more than 85%. The size obtained by the visual algorithm is fitted to the weight of the swimming crab. In order to improve the matching degree of the model, the k-means clustering is introduced to implement a dual-model support vector regression machine(SVR) prediction structure, which uses the differential evolution algorithm to optimize the fitness function of the SVR. In the test process, after normalizing the newly input data, the cluster center to which it belongs is first determined, and then the corresponding SVR model is applied for prediction. The test set results show that the relative error remains below 18%.更多还原