Drone-based apple detection: Finding the depth of apples using YOLOv7 architecture with multi-head attention mechanism

Abstract

The agriculture drones are flocking and hovering over the crop fields to collect data or perform tasks related to crop management. The rise of artificial intelligence and machine learning algorithms paved the path to innovative approaches in agriculture. Apple detection on Apple farms has been a distinguished area of expertise. The apple target recognition algorithms like YOLOv7 gained a lot of reflection and accuracy to identify, recognize and detect the apples. However, the significant problems with the accurate time detection of apples include occlusions, wiring, branching, and overlapping. So, to overcome this problem, a deep learning approach of the Apple detection model is projected to rectify the margin of error with drone-based inference on the live field. Along with the accurate detection of apples, the depth of apples from a drone offers valued data for optimizing harvesting, assessing yield, discovering diseases, handling orchards, and evolving agricultural research in the apple industry. A specific multi-head attention mechanism is applied to the model to capture spatial and channel-wise dependencies concurrently. It can help capture complex interactions between regions and features, improving apple detection accuracy. The designed model is to detect apples in complex backgrounds better. The model identifies minimal objects and enhances the quality of features to achieve accurate bounding boxes, which maximizes detection accuracy. Incorporating a function to evaluate loss further increases the model's accuracy. According to a comparative study, the proposed model using the modified Yolov7 architecture attains a good accuracy of 0.91, 0.96, and 0.92 concerning precision, recall, and F1-score, respectively.

Topics & Concepts