SOD head: A network for locating small fruits from top to bottom in layers of feature maps

Abstract

Although object detection technology has been applied in the field of smart orchards, detecting small fruits in real orchard environments is still a great challenge due to the interference of fruit scale issues. In this study, we propose an effective detection head named SOD Head for detecting small-scale fruits in the early growth stage, aiming to enhance the monitoring of fruit growth in the early stages and achieve intelligent management of orchards. SOD Head firstly utilizes the rich semantic information in the top-level feature map to determine the vague feature position, and mapping downward to the next level, achieving layer-by-layer locating and refinement of feature information. This can avoid missing the features of small fruits that are sparse on the high-resolution feature map and reduce the interference brought by information redundancy to small-scale detection. Secondly, SOD Head performs operation of box relocation to make the prediction of the boundary boxes for small-scale fruits more stable. The experimental results show that SOD Head achieves APs of 29.5% and 39.6% on the datasets of Gold Pear before the thinning stage and MinneApple respectively. Overall, SOD Head not only has a higher detection accuracy on small-scale fruits than other algorithms, but also has good generalization and versatility.