节点文献
基于SSD目标检测框架的乌龟常见病症识别方法
The Method of Tortoise Common Disease Identification Based on SSD Object Detection Framework
【摘要】 为实现对乌龟常见病症的快速检测,以便及时对患病乌龟进行治疗,防止病情加重或传染蔓延,减少乌龟养殖所需要的人力监督,降低养殖过程因监督不当所造成的乌龟因病死亡带来的损失。结合深度神经网络,采用可分卷积代替传统卷积方式,精简SSD300目标检测框架以及特征提取网络结构,以retina_net中提出的Focal Loss作为损失函数训练目标检测模型,并且针对乌龟常见状态的目标尺寸,利用聚类算法k-means得出适合乌龟病症识别的默认框纵横比,使预测框的回归更加精确,最终利用非极大值抑制去除重叠率较大的预测框,得出检测结果。相比于SSD300原模型,该模型参数量从550.1MB减少至18.8MB,参数量缩减共计531.3MB,检测一张图像仅需0.45s,速度提升4.11s,平均查准率为98.22%,仅仅降低0.48%,同时也验证了Focal Loss对于目标检测的提升,采用Focal Loss的模型至少提升2个百分比的平均查准率。该方法能够有效地检测出乌龟的白眼病、中耳炎、腐甲病3种常见病症,在保证精度的同时,大幅提升检测速度,能及时发现具有很强传染性的腐甲病病龟,实施相应的治疗和隔离措施后,可防止腐甲病进一步扩散,极大程度的降低损失。
【Abstract】 In order to reach the quick detection of common diseases of turtles, so as to treat sick turtles in time, prevent the disease aggravation or spread of infection, reduce the human supervision needed for turtle breeding and reduce the loss caused by disease death of turtles due to the lack of supervision in the breeding process, We combined with the deep neural network,and used the separable convolution to replace traditional convolution, simplified SSD300 object detection framework and network structure, took the Focal Loss which was proposed in Retina_net as the Loss function to train the model. In addition, the clustering algorithm k-means is used to obtain the default box aspect ratio suitable for turtle disease identification, so as to make the regression of the prediction box more accurate. Finally, the prediction box with large overlap ratio is eliminated by non-maximum suppression. Compared with the original SSD300 model, the number of model parameters decreased from 550.1 MB to 18.8 MB, and the parameters decreased by 531.3 mb in total. It only took 0.45 s to detect an image, the speed increased by4.11 s, and the average precision was 98.22%, only 0.48% lower. At the same time, the promotion of Focal Loss to object detection was also verified, the average precision of models which take Focal Loss as the loss function was improved at least 2 percentage. This method can effectively detect three kinds of common disease which are white eye disease, otitis media, rotten shell disease, and it can improve the speed of detection greatly and ensure the accuracy at the same time. It can timely find the turtle with rotten nail disease, the implementation of the corresponding treatment and isolation measures, can prevent the further spread of it, greatly reduce the loss.
【Key words】 deep learning; object detection; image processing; SSD; turtle disease recognition;
- 【文献出处】 沈阳农业大学学报 ,Journal of Shenyang Agricultural University , 编辑部邮箱 ,2020年02期
- 【分类号】S947.9;TP391.41
- 【被引频次】1
- 【下载频次】201