Small Target Fruit Detection Under The Unstructured Orchard Environment

The application of advanced theory of deep learning provides new ideas and solutions for scientific and intelligent management of intelligent orchard throughout the cycle.Accurate detection is an effective tools for orchard yield measurement and full time monitoring.However,the detection precision of small target fruit is seriously low due to factors such as the the closeness color between fruit skin and background as well as the small pixel ratio in the unstructured orchard environment.Therefore,exploring the accurate detection of small target fruit has become a hot topic in the field of intelligent agriculture.This study takes early-growth fruits and visionary target fruits of apple orchards as the research objects,and focuses on the efficient and accurate detection of small target fruits of apples.We evaluate our methods on two datasets of GreenApple and MinneApple.The main contents are as follows:(1)To address the imbalance in sample size between target fruit and background,the small target fruit detection model GHFormer-Net is constructed.Feature information is extracted from global receptive field by building the PVTv2-B1 structure;The classification loss function and the boundary box regression loss function are optimized through the gradient harmonizing mechanism.The experimental results show that the average detection precision of GHFormer-Net for small target reaches 46.8% and 41.3% on the two data sets without increasing the computational overhead.(2)To alleviate the problem of imbalance between large and small scale fruit feature information,a small target fruit detection model of balanced feature pyramids network,BFP-Net,is proposed.The imbalance of small-scale target information on different layers of FPN is alleviated by constructing a weight-like feature fusion architecture;The extended feature is introduced and the decoupled-aggregated module is designed in the extended layer of the lowest layer of FPN to supplement the spatial location information;distillation mechanism is introduced to achieve favourable effective transfer of knowledge.Experimental results show that the average precision for small target of BFP-Net reaches 47.0% and 42.2% on the two datasets respectively.(3)To address the issue of insufficient interaction between local and global features,an interactive local and global feature network ILG-Net is designed.By constructing a multi-granularity Transformer based on window to capture global coarse-grained and local fine-grained information;designing a dual-stream feature aggregation block to fuse local and global features;devising a global guidance strategy to guide local information stream and mitigate the edge blur caused by colour.Experimental results show that the average precision for small targets of ILG-Net on the two datasets is 47.8% and 42.4% respectively.In summary,the proposed small target detection models improve the recognition precision of small-scale target fruit in the same colour family in the unstructured apple orchard environment,effectively alleviating the problem of fruit yield measurement of small target fruit in the same color system in the early stage of fruit growth,which can provide technical support for automated detection during the whole apple growing period.