Auto-adjustment label assignment-based convolutional neural network for oriented wheat diseases detection.

• A convolutional neural network-based method is proposed for detecting wheat diseases characterized by arbitrary orientation and significant aspect ratio variations. • An auto-adjustment label assignment scheme based on the similarity of aspect ratio between the sample and object is proposed to assign high-potential positive samples. • A localization potential assessment scheme based on location distance is proposed to evaluate each positive sample. • A wheat disease dataset WDF2023 based on oriented annotation is established. The frequent occurrence of wheat diseases seriously affects the quality and yield of wheat. Thus, the accurate detection of wheat diseases is highly desired in the field of agricultural information. However, for wheat disease with arbitrary-oriented and aspect ratio varies greatly, existing deep learning-based methods adopt defined IoU threshold to assign label and do not consider the differences in localization potential of selected positive samples, resulting in some objects with large aspect ratios could not match enough high potential positive samples. In this paper, we put forward a convolutional neural network-based method for detecting wheat diseases characterized by arbitrary orientation and significant aspect ratio variations. First, we design an auto-adjustment label assignment scheme based on the similarity of aspect ratio between the sample and object to assign high-potential positive samples. Then, a localization potential assessment scheme is proposed to evaluate each positive sample. Finally, we construct a dataset of wheat disease in field (WDF2023) based on oriented annotation. We evaluate the effectiveness of our proposed method and eight oriented object detection detectors. The experimental results showcase that the proposed method attains an mAP of 60.8% and an mRecall of 73.8% on the WDF2023 dataset, surpassing existing advanced oriented object detection detectors. Notably, when contrasted with conventional horizontal object detection detectors, our method demonstrates superior performance in precisely localizing disease regions. [ABSTRACT FROM AUTHOR]