| 多目标鱼体对象提议检测算法研究 |
| |
| 引用本文： | 孙龙清,刘婷,陈帅华,吴雨寒.多目标鱼体对象提议检测算法研究[J].农业机械学报,2019,50(12):260-267. |
| |
| 作者姓名： | 孙龙清 刘婷 陈帅华 吴雨寒 |
| |
| 作者单位： | 中国农业大学,中国农业大学,中国农业大学,中国农业大学 |
| |
| 基金项目： | “十二五”国家科技支撑计划项目（2015BAD17B04-5） |
| |
| 摘 要： | 鱼的行为变化除了可以反映其自身健康状况外，还对分析水质变化具有重要意义，而精确、快速的鱼体目标检测是行为变化分析的基础。针对现有多目标鱼体检测算法存在检测定位精确度低的问题，提出了一种简单、有效的多目标鱼体对象提议检测算法。提取鱼体图像的骨架和边缘信息，制定新的窗口打分策略生成候选窗口，训练PCA卷积核提取鱼体图像前景和背景特征，利用支持向量机（Support vector machine,SVM）识别得到鱼体目标所在的候选窗口，运用非极大值抑制算法剔除冗余窗口完成目标检测。实验表明，基于新的窗口打分策略生成的候选窗口比Edge Boxes算法得到的候选窗口具有更高的召回率，召回率最高可达96.9%，对候选窗口的最高识别准确率可达95.71%。通过本文算法和Edge Boxes-PCANet算法得到的漏检率、误检率和平均检测时间表明，本文算法的综合表现更优，说明本文算法可以高效精确地实现多目标鱼体检测。
|
| 关 键 词： | 鱼体 目标检测 边缘 骨架 支持向量机 |
| 收稿时间： | 2019/8/25 0:00:00 |
Multi-target Fish Detection Algorithm Based on Object Proposals |
| |
| Institution: | China Agricultural University,China Agricultural University,China Agricultural University and China Agricultural University |
| |
| Abstract: | In addition to reflecting its own health status, fish behavioral changes are also important in analyzing water quality. The accurate and rapid fish detection is the basis for behavioral change analysis. In order to solve the problem of low precision in the existing multi target fish detection algorithms, a simple but effective multi target fish detection algorithm was proposed. A new window scoring strategy was created to generate proposal windows by using the skeleton and edge cues of the fish image. The principal component analysis convolution kernels were trained to extract foreground and background features of fish images. The support vector machine was used to classify proposal windows to obtain windows where fish were located, and the non maximum suppression algorithm was used to eliminate redundant windows to complete the object detection. Experiments showed that the proposed algorithm based on the new window scoring strategy had a higher recall rate than the Edge Boxes algorithm, and the recall rate was up to 96.9% under the fixed proposals. The highest classification accuracy of proposal windows can reach 95.71%. By analyzing the missed detection rate, false detection rate and average detection time of the algorithm and Edge Boxes-PCANet, the overall performance of the algorithm was optimal. Using this detection algorithm, the multi target fish detection can be achieved efficiently and accurately. |
| |
| Keywords: | fish object detection edge skeleton support vector machine |
|
| 点击此处可从《农业机械学报》浏览原始摘要信息 |
| 点击此处可从《农业机械学报》下载免费的PDF全文 |
|
