This paper describes a study on the development of an artificial intelligence system for a robot following a strawberry picking robot. In order to pick a strawberry, you first need to judge strawberry ripe well or not to avoid picking strawberries that are not ripe. To do this, we need to create two...
This paper describes a study on the development of an artificial intelligence system for a robot following a strawberry picking robot. In order to pick a strawberry, you first need to judge strawberry ripe well or not to avoid picking strawberries that are not ripe. To do this, we need to create two deep-learning systems that determine the ripeness of strawberries and find the coordinates of vertices for picking strawberries. However, since creating two networks is inefficient and requires high parameters and a lot of time, this paper proposed a network that performs both of these simultaneously.
Semantic Segmentation was used as a method for detecting the ripeness of Strawberries. There are three classes in total: unripe strawberries, medium ripe strawberries, and ripe strawberries. As a method of finding the coordinates of the vertices of the strawberry, the method of finding the Keypoints segmentation was used. I wanted a circle with a radius of 15pixels where I wanted it, and the value of this circel was 255, and the background was 0. At this time, as the distance from the center of the circle became smaller, the value became smaller to help learning. Gaussian filter was used as a method for this.
In this paper, a multi-path Convolution method is proposed. This is a method to determine that the values of the parameters according to the number of the previous feature map and the next feature map are unnecessarily large when convolution, and divide them into two paths to extract only the feature maps with half values and later extract the two feature maps. Ths same number of feature maps. The same number of feature maps were extracted for each convolution while reducing the parameters by convergence.
As a model for two segments, a multi-path U-net model was proposed. This is a U-net using the Multi-path Convolution proposed in this paper, and it is divided into two paths at the last stage, and two segments are performed through the feature map obtained after Multi-path Convolution. Through this, Semantic Segmentation to know the ripeness of strawberries obtained mIoU value of 92% and Keypoints Segmentation to obtain coordinates of vertices obtained mIoU value of 74%. This is thought to be a great help for the robot to develop an automatic unmanned farm system for picking strawberries.
This paper describes a study on the development of an artificial intelligence system for a robot following a strawberry picking robot. In order to pick a strawberry, you first need to judge strawberry ripe well or not to avoid picking strawberries that are not ripe. To do this, we need to create two deep-learning systems that determine the ripeness of strawberries and find the coordinates of vertices for picking strawberries. However, since creating two networks is inefficient and requires high parameters and a lot of time, this paper proposed a network that performs both of these simultaneously.
Semantic Segmentation was used as a method for detecting the ripeness of Strawberries. There are three classes in total: unripe strawberries, medium ripe strawberries, and ripe strawberries. As a method of finding the coordinates of the vertices of the strawberry, the method of finding the Keypoints segmentation was used. I wanted a circle with a radius of 15pixels where I wanted it, and the value of this circel was 255, and the background was 0. At this time, as the distance from the center of the circle became smaller, the value became smaller to help learning. Gaussian filter was used as a method for this.
In this paper, a multi-path Convolution method is proposed. This is a method to determine that the values of the parameters according to the number of the previous feature map and the next feature map are unnecessarily large when convolution, and divide them into two paths to extract only the feature maps with half values and later extract the two feature maps. Ths same number of feature maps. The same number of feature maps were extracted for each convolution while reducing the parameters by convergence.
As a model for two segments, a multi-path U-net model was proposed. This is a U-net using the Multi-path Convolution proposed in this paper, and it is divided into two paths at the last stage, and two segments are performed through the feature map obtained after Multi-path Convolution. Through this, Semantic Segmentation to know the ripeness of strawberries obtained mIoU value of 92% and Keypoints Segmentation to obtain coordinates of vertices obtained mIoU value of 74%. This is thought to be a great help for the robot to develop an automatic unmanned farm system for picking strawberries.
※ AI-Helper는 부적절한 답변을 할 수 있습니다.