For practitioners, it is very crucial to realize accurate and automatic vision-based quality identification of Longjing tea. Due to the high similarity between classes, the classification accuracy of traditional image processing combined with machine learning algorithm is not satisfactory. High-performance deep learning methods require large amounts of annotated data, but collecting and labeling massive amounts of data is very time consuming and monotonous. To gain as much useful knowledge as possible from related tasks, an instance-based deep transfer learning method for the quality identification of Longjing tea is proposed. The method mainly consists of two steps: (i) The MobileNet V2 model is trained using the hybrid training dataset containing all labeled samples from source and target domains. The trained MobileNet V2 model is used as a feature extractor, and (ii) the extracted features are input into the proposed multiclass TrAdaBoost algorithm for training and identification. Longjing tea images from three geographical origins, West Lake, Qiantang, and Yuezhou, are collected, and the tea from each geographical origin contains four grades. The Longjing tea from West Lake is regarded as the source domain, which contains more labeled samples. The Longjing tea from the other two geographical origins contains only limited labeled samples, which are regarded as the target domain. Comparative experimental results show that the method with the best performance is the MobileNet V2 feature extractor trained with a hybrid training dataset combined with multiclass TrAdaBoost with linear support vector machine (SVM). The overall Longjing tea quality identification accuracy is 93.6% and 91.5% on the two target domain datasets, respectively. The proposed method can achieve accurate quality identification of Longjing tea with limited samples. It can provide some heuristics for designing image-based tea quality identification systems.

Image-based fruit classification offers many useful applications in industrial production and daily life, such as self-checkout in the supermarket, automatic fruit sorting and dietary guidance. However, fruit classification task will have different data distributions due to different application scenarios. One feasible solution to solve this problem is to use domain adaptation that adapts knowledge from the original training data (source domain) to the new testing data (target domain). In this paper, we propose a novel deep learning-based unsupervised domain adaptation method for cross-domain fruit classification. A hybrid attention module is proposed and added to MobileNet V3 to construct the HAM-MobileNet that can suppress the impact of complex backgrounds and extract more discriminative features. A hybrid loss function combining subdomain alignment and implicit distribution metrics is used to reduce domain discrepancy during model training and improve model classification performance. Two fruit classification datasets covering several domains are established to simulate common industrial and daily life application scenarios. We validate the proposed method on our constructed grape classification dataset and general fruit classification dataset. The experimental results show that the proposed method achieves an average accuracy of 95.0% and 93.2% on the two datasets, respectively. The classification model after domain adaptation can well overcome the domain discrepancy brought by different fruit classification scenarios. Meanwhile, the proposed datasets and method can serve as a benchmark for future cross-domain fruit classification research.

No abstract

No abstract

Catching high-speed targets in the flight is a complex and typical highly dynamic task. However, existing methods require manual setting of catching height or time, resulting in lacks of adaptability and flexibility and cannot deal with multiple targets. To bridge this gap, we propose a planning-with-decision scheme called Catch Planner. For sequential decision making, a lightweight policy search method based on deep reinforcement learning is proposed. It is jointly trained with the motion planning and decoupled from physics to speed up training. For motion planning, we propose a trajectory optimization method that jointly optimizes the highly coupled catching time and terminal state. The core is the flexible-terminal constraint transcription. It converts the three unique constraints of catching into differentiable metrics, including equality constraints for terminal position and time, and inequality constraints that enable reasonable terminal position offset and attitude relaxation. In addition, sparse parameterization based on MINCO class considers both dynamic feasibility and collision avoidance constraints. As a result, a generally constrained quadrotor planning problem is transformed into an unconstrained optimization that can be solved reliably and efficiently. We also propose an online iterative optimization method for predicting differentiable trajectories of targets. Catch Planner provides a new paradigm for the combination of learning and planning, where all algorithms can be run in real time onboard at 100hz. Extensive experiments are carried out in real-world and simulated scenes to verify the robustness and expansibility when facing a variety of high-speed flying targets.