Machine Vision for Strawberry Detection

Abstract. Strawberry growers rely on accurate yield forecasts to reduce spoilage, minimize harvest costs, and formulate contracts. Currently, strawberry yields are estimated by manually counting a small section of berries and extrapolating to forecast the yield for the entire field. These estimates have high variance and are labor-intensive. This paper proposes a fully-automated strawberry detection method that can eliminate much of the manual effort required for yield estimation. The approach uses machine vision approach for counting strawberries at various stages of growth in images. These counts can then be used to forecast yield in the future. Images can be gathered from drones or using cameras attached to existing tractors. These images can then be processed to determine how many strawberries are present at each stage of growth. For our experiments, a consumer-grade camera was used to take photographs of strawberries in a field on the California State University of San Luis Obispo. The challenging dataset includes more than 400 images from two different perspectives where many berries are severely occluded. The Overfeat neural network model was trained to detect berries. The models achieved a mean average precision (MAP) of .79 across two perspectives and three stages of strawberry growth. A major advantage of the system is that there is no need for fine-tuning parameters or customization, and the model performed well with just a few hundred labeled training examples. This machine vision technique can be used as an input to an expert system to make strawberry yield forecasts in the future.