A greenhouse strawberry seedbed image stitching method based on seedbed plane alignment and efficient point-line matching

• Seedbed Plane Alignment and efficient Point-Line matching improve stitching quality. • Optimize the registration matrix using the camera’s equal-interval imaging characteristics. • Image Distortion Degree (IDD) is proposed to assess the naturalness of stitching. • Our method preserves the seedbed’s geometric structure and maintains overall naturalness. • This study provides high-quality panoramic images for monitoring strawberry growth. In the context of factory-based cultivation of strawberry seedlings in greenhouses, acquiring panoramic images of the seedbed is essential for monitoring the overall growth of the strawberry seedlings, including assessing the uniformity of growth and detecting the presence of pests and diseases. This paper presents a novel greenhouse seedbed image stitching method based on seedbed plane alignment and efficient point-line matching, using a rail-based inspection device mounted above the seedbed to capture sequential images covering the seedbed area, in order to obtain high-quality panoramic images of the strawberry seedlings. First, since the strawberry seedlings are located within the seedbed area, the Depth-Anything model is utilized to extract the seedbed plane, allowing the registration algorithm to focus on the precise alignment of the seedbed region. Secondly, to fully leverage the geometric structure in the strawberry seedbed images, a local registration method GlueStick based on point-line matching is applied to match feature points between overlapping seedbed images, significantly reducing the number of feature points while effectively enhancing matching accuracy. Finally, exploiting the equidistant imaging characteristics of the rail-based imaging device, a homography matrix optimization method is proposed, effectively mitigating the impact of a small number of inaccurate local matches on the global stitching performance. Comprehensive experiments, encompassing both subjective (qualitative scoring) and objective (RMSE, Image Distortion Degree) evaluations, conducted on the constructed strawberry seedbed image dataset, demonstrate that the proposed method achieves precise alignment of the seedbed and effectively preserves the overall naturalness, outperforming representative image stitching methods. The proposed method delivers high-quality panoramic images for seedbed monitoring, offering substantial support for precise monitoring of greenhouse crops, and provides valuable references for panoramic stitching methods of other greenhouse crops.