A comparative study on deep learning algorithms performance on flax crop boll-counting from crop RGB images

Abstract. In crop breeding, being able to estimate yield from a variety of genotypes prior to harvest can save plant breeders time and resources. One important trait for yield estimation, for crops such as soybean and flax, is the number of pods per plant. Manual counting of pods can be a labor-intensive, inaccurate, and time-consuming task. That is especially true for flax, which has many small pods (bolls) per plant. For such situations, non-destructive machine vision techniques applied to field-collected images (e.g., RGB cameras) offer the possibility of a faster and more accurate bolls counting, which in turn can be used for yield estimation prior to harvest. Deep learning (DL) frameworks could assist to build a boll counting model based on flax RGB images. Although DL frameworks have been extensively used to estimate the yield of different crops, we are not aware of any study regarding the application of DL algorithms in counting bolls on flax crops. The objectives of this study were 1) to test different DL algorithms including YOLOv7, SSD ResNet152 V1 FPN, EfficientDet D1 640x640, and Faster R-CNN Inception ResNet V2; and 2) to compare their performance to find the best architecture for bolls counting by testing and evaluating the model using COCO-evaluation metrics. Flax images were acquired by a Google Pixel 4 smartphone camera (12.2 megapixels [MP]) and a Sony handycam FDR-ax53 (16.6 MP) in Casselton, ND, in the summer 2022, between green capsule and seed ripe growth stages. A total number of 1000 images were used for this experiment and to enhance the dataset, generalize the trained model, and improve predictions, various data augmentation techniques were applied to the image dataset. This study demonstrated that among the tested DL algorithm, the YOLOv7 model has the best potential of estimating flax boll-counts with acceptable accuracy. That model can significantly reduce the time and effort required to measure the number of bolls in flax, which can in turn increase efficiencies in a flax breeding program.