3D Reconstruction Method for Strawberry Plants Based on 3D Gaussian Splatting and Edge Detection

During the 3D reconstruction of strawberry plants, methods based on 3D Gaussian Splatting (3DGS) face significant challenges due to motion-induced image blur. Such blurring substantially reduces the feature matching accuracy in Structure from Motion (SfM) algorithms and compromises the reliability of camera pose estimation, thereby degrading the quality of subsequent 3DGS reconstruction. This ultimately manifests as geometric distortion and loss of texture details in the reconstructed models. The issue is particularly severe on the surface of strawberry fruits: under blurred image conditions, point cloud registration fails, resulting in the loss of high-frequency details in the high-density achene regions, which blurs seed contours and degrades reconstruction accuracy. To address this technical bottleneck, this study proposes an optimized reconstruction scheme integrated with 3DGS. By incorporating the Canny edge detection algorithm to filter h i gh-quality i n put i m ages, t h e a c curacy of the reconstructed model is significantly improved. The optimized approach achieves remarkable results on the strawberry plant dataset: the average Peak Signal-To-Noise Ratio (PSNR) of the 3DGS model reaches 35.99, representing a 15.2% improvement over the baseline 3DGS. The morphology of high-density achenes on the fruit surface is clearly distinguishable, supporting the accurate monitoring of phenotypic parameters in strawberry plants.