Search

Although plant disease recognition has witnessed a significant improvement with deep learning in recent years, a common observation is that current deep learning methods with decent performance tend to suffer in real-world applications. We argue that this illusion essentially comes from the fact that current plant disease recognition datasets cater to deep learning methods and are far from real scenarios. Mitigating this illusion fundamentally requires an interdisciplinary perspective from both plant disease and deep learning, and a core question arises. What are the characteristics of a desired dataset? This paper aims to provide a perspective on this question. First, we present a taxonomy to describe potential plant disease datasets, which provides a bridge between the two research fields. We then give several directions for making future datasets, such as creating challenge-oriented datasets. We believe that our paper will contribute to creating datasets that can help achieve the ultimate objective of deploying deep learning in real-world plant disease recognition applications. To facilitate the community, our project is publicly available at https://github.com/xml94/PPDRD with the information of relevant public datasets.

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.

We propose an enhanced class-specific spatial normalization, a simple yet effective layer to generate a photorealistic image given a spatial-class map. Under the assumption that pixels belonging to the same class share the same distribution in the feature space, we intuitively split an image into classes according to the map. By learning the class-specific distributions, our generator can distinguish one class from other classes. Further, our spatial normalization combines the spatial-class map and the class-specific distributions, by which our generator can produce instances in the desired locations. We apply the proposed normalization not only in semantic image generation but also in object transfiguration. The experimental results demonstrate that the spatial-class map can be efficiently utilized with our proposed method, which results in competing performances with much fewer parameters.

Object detection models have become the current tool of choice for plant disease detection in precision agriculture. Most existing research improved the performance by ameliorating networks and optimizing the loss function. However, because of the vast influence of data annotation quality and the cost of annotation, the data-centric part of a project also needs more investigation. We should further consider the relationship between data annotation strategies, annotation quality, and the model’s performance. In this paper, a systematic strategy with four annotation strategies for plant disease detection is proposed: local, semi-global, global, and symptom-adaptive annotation. Labels with different annotation strategies will result in distinct models’ performance, and their contrasts are remarkable. An interpretability study of the annotation strategy is conducted by using class activation maps. In addition, we define five types of inconsistencies in the annotation process and investigate the severity of the impact of inconsistent labels on model’s performance. Finally, we discuss the problem of label inconsistency during data augmentation. Overall, this data-centric quantitative analysis helps us to understand the significance of annotation strategies, which provides practitioners a way to obtain higher performance and reduce annotation costs on plant disease detection. Our work encourages researchers to pay more attention to annotation consistency and the essential issues of annotation strategy. The code will be released at: https://github.com/JiuqingDong/PlantDiseaseDetection_Yolov5 .

Deep learning has witnessed a significant improvement in recent years to recognize plant diseases by observing their corresponding images. To have a decent performance, current deep learning models tend to require a large-scale dataset. However, collecting a dataset is expensive and time-consuming. Hence, the limited data is one of the main challenges to getting the desired recognition accuracy. Although transfer learning is heavily discussed and verified as an effective and efficient method to mitigate the challenge, most proposed methods focus on one or two specific datasets. In this paper, we propose a novel transfer learning strategy to have a high performance for versatile plant disease recognition, on multiple plant disease datasets. Our transfer learning strategy differs from the current popular one due to the following factors. First, PlantCLEF2022, a large-scale dataset related to plants with 2,885,052 images and 80,000 classes, is utilized to pre-train a model. Second, we adopt a vision transformer (ViT) model, instead of a convolution neural network. Third, the ViT model undergoes transfer learning twice to save computations. Fourth, the model is first pre-trained in ImageNet with a self-supervised loss function and with a supervised loss function in PlantCLEF2022. We apply our method to 12 plant disease datasets and the experimental results suggest that our method surpasses the popular one by a clear margin for different dataset settings. Specifically, our proposed method achieves a mean testing accuracy of 86.29over the 12 datasets in a 20-shot case, 12.76 higher than the current state-of-the-art method’s accuracy of 73.53. Furthermore, our method outperforms other methods in one plant growth stage prediction and the one weed recognition dataset. To encourage the community and related applications, we have made public our codes and pre-trained model1.

Predicting plant growth is a fundamental challenge that can be employed to analyze plants and further make decisions to have healthy plants with high yields. Deep learning has recently been showing its potential to address this challenge in recent years, however, there are still two issues. First, image-based plant growth prediction is currently taken either from time series or image generation viewpoints, resulting in a flexible learning framework and clear predictions, respectively. Second, deep learning-based algorithms are notorious to require a large-scale dataset to obtain a competing performance but collecting enough data is time-consuming and expensive. To address the issues, we consider the plant growth prediction from both viewpoints with two new time-series data augmentation algorithms. To be more specific, we raise a new framework with a length-changeable time-series processing unit to generate images flexibly. A generative adversarial loss is utilized to optimize our model to obtain high-quality images. Furthermore, we first recognize three key points to perform time-series data augmentation and then put forward T-Mixup and T-Copy-Paste. T-Mixup fuses images from a different time pixel-wise while T-Copy-Paste makes new time-series images with a different background by reusing individual leaves extracted from the existing dataset. We perform our method in a public dataset and achieve superior results, such as the generated RGB images and instance masks securing an average PSNR of 27.53 and 27.62, respectively, compared to the previously best 26.55 and 26.92.

Instance-level image translation aims to only translate instance of interest and can be operated more finely and flexibly than object-level and holistic-level image translation. However, current algorithms are not suitable to do it since they employ a holistic or object level’s discriminator that tends to change the whole image or all instances. To address the issue, we propose a simple yet effective local discriminator, in which the input image is split into two parts, region of interest (ROI) and background. Instance mask is employed to align the ROI and the background is design to be random in a prior distribution to mitigate a divergence between the ROI and the background. In this way, we obtain translated instance with decent margins without artifacts as current algorithms get. Moreover we propose a new architecture to simultaneously realize versatile instance-level image translation. Experimental results prove that our proposed algorithm outperforms the state-of-the-art in position accuracy and background retainment by a clear margin.

Deep learning shows its advantages and potentials in plant disease recognition and has witnessed a profound development in recent years. To obtain a competing performance with a deep learning algorithm, enough amount of annotated data is requested but in the natural world, scarce or imbalanced data are common, and annotated data is expensive or hard to collect. Data augmentation, aiming to create variations for training data, has shown its power for this issue. But there are still two challenges: creating more desirable variations for scarce and imbalanced data, and designing a data augmentation to ease object detection and instance segmentation. First, current algorithms made variations only inside one specific class, but more desirable variations can further promote performance. To address this issue, we propose a novel data augmentation paradigm that can adapt variations from one class to another. In the novel paradigm, an image in the source domain is translated into the target domain, while the variations unrelated to the domain are maintained. For example, an image with a healthy tomato leaf is translated into a powdery mildew image but the variations of the healthy leaf are maintained and transferred into the powdery mildew class, such as types of tomato leaf, sizes, and viewpoints. Second, current data augmentation is suitable to promote the image classification model but may not be appropriate to alleviate object detection and instance segmentation model, mainly because the necessary annotations can not be obtained. In this study, we leverage a prior mask as input to tell the area we are interested in and reuse the original annotations. In this way, our proposed algorithm can be utilized to do the three tasks simultaneously. Further, We collect 1,258 images of tomato leaves with 1,429 instance segmentation annotations as there is more than one instance in one single image, including five diseases and healthy leaves. Extensive experimental results on the collected images validate that our new data augmentation algorithm makes useful variations and contributes to improving performance for diverse deep learning-based methods.

To balance the convergence speed and the solution's diversity in the large-scale travel salesman problem (TSP), this paper proposes a new heuristic communication heterogeneous dual population ant colony optimization (HHACO). First, the main characteristics of HHACO are the heuristic communication and the two heterogeneous ant colonies. Heuristic communication, an indirect communication strategy, helps improve the deviation of solution. Heterogeneous ant colonies are beneficial to balance the convergence speed and the diversity of solution, in which one ant colony is in charge of solution's diversity and the another one in charge of convergence speed inspiring from nature evolution with self-adaptive ability. Besides, this paper takes advantage of orthogonal test to discuss the parameters in HHACO algorithm and a better parameters' set is obtained. Then, HHACO algorithm is applied to solve TSP, and meanwhile characteristics of different ant colonies in HHACO are discussed. Finally, HHACO is compared with the other dual colonies algorithms and several classic ant colony optimization algorithms, and results suggest that the HHACO has a better performance in the large-scale problem.