Search

Plant disease detection has made significant strides thanks to the emergence of deep learning. However, existing methods have been limited to closed-set and static learning settings, where models are trained using a specific dataset. This confinement restricts the model’s adaptability when encountering samples from unseen disease categories. Additionally, there is a challenge of knowledge degradation for these static learning settings, as the acquisition of new knowledge tends to overwrite the old when learning new categories. To overcome these limitations, this study introduces a novel paradigm for plant disease detection called open-world setting. Our approach can infer disease categories that have never been seen during the model training phase and gradually learn these unseen diseases through dynamic knowledge updates in the next training phase. Specifically, we utilize a well-trained unknown-aware region proposal network to generate pseudo-labels for unknown diseases during training and employ a class-agnostic classifier to enhance the recall rate for unknown diseases. Besides, we employ a sample replay strategy to maintain recognition ability for previously learned classes. Extensive experimental evaluation and ablation studies investigate the efficacy of our method in detecting old and unknown classes. Remarkably, our method demonstrates robust generalization ability even in cross-species disease detection experiments. Overall, this open-world and dynamically updated detection method shows promising potential to become the future paradigm for plant disease detection. We discuss open issues including classification and localization, and propose promising approaches to address them. We encourage further research in the community to tackle the crucial challenges in open-world plant disease detection. The code will be released at https://github.com/JiuqingDong/OWPDD.

Previous work on plant disease detection demonstrated that object detectors generally suffer from degraded training data, and annotations with noise may cause the training task to fail. Well-annotated datasets are therefore crucial to build a robust detector. However, a good label set generally requires much expert knowledge and meticulous work, which is expensive and time-consuming. This paper aims to learn robust feature representations with inaccurate bounding boxes, thereby reducing the model requirements for annotation quality. Specifically, we analyze the distribution of noisy annotations in the real world. A teacher-student learning paradigm is proposed to correct inaccurate bounding boxes. The teacher model is used to rectify the degraded bounding boxes, and the student model extracts more robust feature representations from the corrected bounding boxes. Furthermore, the method can be easily generalized to semi-supervised learning paradigms and auto-labeling techniques. Experimental results show that applying our method to the Faster-RCNN detector achieves a 26% performance improvement on the noisy dataset. Besides, our method achieves approximately 75% of the performance of a fully supervised object detector when 1% of the labels are available. Overall, this work provides a robust solution to real-world location noise. It alleviates the challenges posed by noisy data to precision agriculture, optimizes data labeling technology, and encourages practitioners to further investigate plant disease detection and intelligent agriculture at a lower cost. The code will be released at https://github.com/JiuqingDong/TS_OAMIL-for-Plant-disease-detection.

Recent advancements in deep learning have brought significant improvements to plant disease recognition. However, achieving satisfactory performance often requires high-quality training datasets, which are challenging and expensive to collect. Consequently, the practical application of current deep learning–based methods in real-world scenarios is hindered by the scarcity of high-quality datasets. In this paper, we argue that embracing poor datasets is viable and aims to explicitly define the challenges associated with using these datasets. To delve into this topic, we analyze the characteristics of high-quality datasets, namely, large-scale images and desired annotation, and contrast them with the limited and imperfect nature of poor datasets. Challenges arise when the training datasets deviate from these characteristics. To provide a comprehensive understanding, we propose a novel and informative taxonomy that categorizes these challenges. Furthermore, we offer a brief overview of existing studies and approaches that address these challenges. We point out that our paper sheds light on the importance of embracing poor datasets, enhances the understanding of the associated challenges, and contributes to the ambitious objective of deploying deep learning in real-world applications. To facilitate the progress, we finally describe several outstanding questions and point out potential future directions. Although our primary focus is on plant disease recognition, we emphasize that the principles of embracing and analyzing poor datasets are applicable to a wider range of domains, including agriculture. Our project is public available at https://github.com/xml94/EmbracingLimitedImperfectTrainingDatasets.

Recent developments in deep learning-based automatic weeding systems have shown promise for unmanned weed eradication. However, accurately distinguishing between crops and weeds in varying field conditions remains a challenge for these systems, as performance deteriorates when applied to new or different fields due to insignificant changes in low-level statistics and a significant gap between training and test data distributions. In this study, we propose an approach based on unsupervised domain adaptation to improve crop-weed recognition in new, unseen fields. Our system addresses this issue by learning to ignore insignificant changes in low-level statistics that cause a decline in performance when applied to new data. The proposed network includes a segmentation module that produces segmentation maps using labeled (training field) data while also minimizing entropy using unlabeled (test field) data simultaneously, and a discriminator module that maximizes the confusion between extracted features from the training and test farm samples. This module uses adversarial optimization to make the segmentation network invariant to changes in the field environment. We evaluated the proposed approach on four different unseen (test) fields and found consistent improvements in performance. These results suggest that the proposed approach can effectively handle changes in new field environments during real field inference.

Plant phenotyping is a critical field in agriculture, aiming to understand crop growth under specific conditions. Recent research uses images to describe plant characteristics by detecting visual information within organs such as leaves, flowers, stems, and fruits. However, processing data in real field conditions, with challenges such as image blurring and occlusion, requires improvement. This paper proposes a deep learning-based approach for leaf instance segmentation with a local refinement mechanism to enhance performance in cluttered backgrounds. The refinement mechanism employs Gaussian low-pass and High-boost filters to enhance target instances and can be applied to the training or testing dataset. An instance segmentation architecture generates segmented masks and detected areas, facilitating the derivation of phenotypic information, such as leaf count and size. Experimental results on a tomato leaf dataset demonstrate the system’s accuracy in segmenting target leaves despite complex backgrounds. The investigation of the refinement mechanism with different kernel sizes reveals that larger kernel sizes benefit the system’s ability to generate more leaf instances when using a High-boost filter, while prediction performance decays with larger Gaussian low-pass filter kernel sizes. This research addresses challenges in real greenhouse scenarios and enables automatic recognition of phenotypic data for smart agriculture. The proposed approach has the potential to enhance agricultural practices, ultimately leading to improved crop yields and productivity.

The emergence of deep neural networks has allowed the development of fully automated and efficient diagnostic systems for plant disease and pest phenotyping. Although previous approaches have proven to be promising, they are limited, especially in real-life scenarios, to properly diagnose and characterize the problem. In this work, we propose a framework which besides recognizing and localizing various plant abnormalities also informs the user about the severity of the diseases infecting the plant. By taking a single image as input, our algorithm is able to generate detailed descriptive phrases (user-defined) that display the location, severity stage, and visual attributes of all the abnormalities that are present in the image. Our framework is composed of three main components. One of them is a detector that accurately and efficiently recognizes and localizes the abnormalities in plants by extracting region-based anomaly features using a deep neural network-based feature extractor. The second one is an encoder–decoder network that performs pixel-level analysis to generate abnormality-specific severity levels. Lastly is an integration unit which aggregates the information of these units and assigns unique IDs to all the detected anomaly instances, thus generating descriptive sentences describing the location, severity, and class of anomalies infecting plants. We discuss two possible ways of utilizing the abovementioned units in a single framework. We evaluate and analyze the efficacy of both approaches on newly constructed diverse paprika disease and pest recognition datasets, comprising six anomaly categories along with 11 different severity levels. Our algorithm achieves mean average precision of 91.7% for the abnormality detection task and a mean panoptic quality score of 70.78% for severity level prediction. Our algorithm provides a practical and cost-efficient solution to farmers that facilitates proper handling of crops.

For continual learning in the process of plant disease recognition it is necessary to first distinguish between unknown diseases from those of known diseases. This paper deals with two different but related deep learning techniques for the detection of unknown plant diseases; Open Set Recognition (OSR) and Out-of-Distribution (OoD) detection. Despite the significant progress in OSR, it is still premature to apply it to fine-grained recognition tasks without outlier exposure that a certain part of OoD data (also called known unknowns) are prepared for training. On the other hand, OoD detection requires intentionally prepared outlier data during training. This paper analyzes two-head network included in OoD detection models, and semi-supervised OpenMatch associated with OSR technology, which explicitly and implicitly assume outlier exposure, respectively. For the experiment, we built an image dataset of eight strawberry diseases. In general, a two-head network and OpenMatch cannot be compared due to different training settings. In our experiment, we changed their training procedures to make them similar for comparison and show that modified training procedures resulted in reasonable performance, including more than 90% accuracy for strawberry disease classification as well as detection of unknown diseases. Accurate detection of unknown diseases is an important prerequisite for continued learning.

Autonomous harvesters can be used for the timely cultivation of high-value crops such as strawberries, where the robots have the capability to identify ripe and unripe crops. However, the real-time segmentation of strawberries in an unbridled farming environment is a challenging task due to fruit occlusion by multiple trusses, stems, and leaves. In this work, we propose a possible solution by constructing a dynamic feature selection mechanism for convolutional neural networks (CNN). The proposed building block namely a dense attention module (DAM) controls the flow of information between the convolutional encoder and decoder. DAM enables hierarchical adaptive feature fusion by exploiting both inter-channel and intra-channel relationships and can be easily integrated into any existing CNN to obtain category-specific feature maps. We validate our attention module through extensive ablation experiments. In addition, a dataset is collected from different strawberry farms and divided into four classes corresponding to different maturity levels of fruits and one is devoted to background. Quantitative analysis of the proposed method showed a 4.1% and 2.32% increase in mean intersection over union, over existing state-of-the-art semantic segmentation models and other attention modules respectively, while simultaneously retaining a processing speed of 53 frames per second.

Weeds in agricultural farms are aggressive growers which compete for nutrition and other resources with the crop and reduce production. The increasing use of chemicals to control them has inadvertent consequences to the human health and the environment. In this work, a novel neural network training method combining semantic graphics for data annotation and an advanced encoder–decoder network for (a) automatic crop line detection and (b) weed (wild millet) detection in paddy fields is proposed. The detected crop lines act as a guiding line for an autonomous weeding robot for inter-row weeding, whereas the detection of weeds enables autonomous intra-row weeding. The proposed data annotation method, semantic graphics, is intuitive, and the desired targets can be annotated easily with minimal labor. Also, the proposed “extended skip network” is an improved deep convolutional encoder–decoder neural network for efficient learning of semantic graphics. Quantitative evaluations of the proposed method demonstrated an increment of 6.29% and 6.14% in mean intersection over union (mIoU), over the baseline network on the task of paddy line detection and wild millet detection, respectively. The proposed method also leads to a 3.56% increment in mIoU and a significantly higher recall compared to a popular bounding box-based object detection approach on the task of wild–millet detection.

Recent developments in deep learning-based automatic weeding systems have shown promise for unmannedweed eradication. However, accurately distinguishing between crops and weeds in varying field conditions remains a challenge for these systems, as performance deteriorates when applied to new or different fields due to insignificant changes in low-level statistics and a significant gap between training and test data distributions. In this study, we propose an approach based on unsupervised domain adaptation to improve crop-weed recognition in new, unseen fields. Our system addresses this issue by learning to ignore insignificant changes in low-level statistics that cause a decline in performance when applied to new data. The proposed network includes a segmentation module that produces segmentation maps using labeled (training field) data while also minimizing entropy using unlabeled (test field) data simultaneously, and a discriminator module that maximizes the confusion between extracted features from the training and test farm samples. This module uses adversarial optimization to make the segmentation network invariant to changes in the field environment. We evaluated the proposed approach on four different unseen (test) fields and found consistent improvements in performance. These results suggest that the proposed approach can effectively handle changes in new field environments during real field inference. [ABSTRACT FROM AUTHOR]