Image recognition Technology in Agricultural Vertical Field based on CNN and pre-training Model
Download as PDF
DOI: 10.23977/iset2021.034
Author(s)
Xu Bailing, Zhang Wanli, Ma Xiaoyu
Corresponding Author
Xu Bailing
ABSTRACT
In order to realize the accurate identification of cow individuals in the complex farm environment, the SSD algorithm is improved to solve the problem that the (single shot multibox detector) algorithm is not effective in detecting overlapping objects. First of all, the feature fusion of different feature images can make different feature images complement each other and improve the detection effect of overlapping objects; then, remove the Conv4_3 layer from the network and increase the number of candidate boxes of other feature images, which can not only ensure the real-time performance of the algorithm but also improve the detection accuracy; finally, the transfer learning method is introduced to improve the average accuracy of the algorithm. The experimental results show that: compared with the traditional SSD algorithm, the improved SSD algorithm improves the average accuracy AP (average precision) by 4.32% while satisfying the real-time detection; after migration, the improved SSD algorithm AP increases by 3.85%.
KEYWORDS
Artificial intelligence, target recognition, convolution neural network, feature fusion, transfer learning