Plant disease detection has an inestimable effect on plant cultivation. Accurate detection of plant disease can control the spread of disease early and prevent unnecessary loss. Strawberry verticillium wilt is a soil-borne, multi-symptomatic disease. To detect strawberry verticillium wilt accurately, we first propose a disease detection network based on Faster R-CNN and multi-task learning to detect strawberry verticillium wilt. Then, the strawberry verticillium wilt detection network (SVWDN), which uses attention mechanisms in the feature extraction of the disease detection network, is proposed. SVWDN detects verticillium wilt according to the symptoms of detected plant components (i.e.,young leaves and petioles). Compared with other existing methods for detecting disease from the whole plant appearance, the SVWDN automatically classifies the petioles and young leaves while determining whether the strawberry has verticillium wilt. To provide a dataset for evaluating and testing our method, we construct a large dataset that contains 3,531 images with 4 categories (Healthy_leaf, Healthy_petiole, Verticillium_leaf and Verticillium_petiole). Each image also has a label to indicate whether the strawberry is suffering from verticillium wilt. With the proposed strawberry verticillium wilt detection network, we achieved a mAP of 77.54% on object detection of 4 categories and 99.95% accuracy for strawberry verticillium wilt detection.

Synthetic Aperture Radar (SAR) has the characteristics of all-weather and all-time operation, which can achieve uninterrupted detection of targets on the sea surface. Currently, small-sized ship targets in SAR images are difficult to detect in complex backgrounds due to limited pixel information, unclear azimuth information, and weak signals after imaging. This makes it challenging to detect small-scale ship targets in SAR images. In this paper, we proposed an Enhanced Cascade R-CNN algorithm for detecting small-sized ship targets in complex backgrounds of SAR images. To enhance the multi-scale expression ability of the network, we introduce Res2Net with richer multi-scale information and establish a spatial enhancement module to increase the weight of the ship target in the aspect map. Additionally, a bidirectional feature pyramid structure is constructed to fuse the feature maps output at numerous stages, making the semantic information contained in the feature maps more abundant. To improve the accuracy of the target boundary in dense areas, we introduce a generalized focal loss function and improve the output layer prediction network. Experiments conducted on the SAR-Ship-Dataset show that our algorithm achieves precision, recall, F1, and mAP of 92.6%, 92.4%, 92.8%, and 92.5%, respectively. The proposed method exhibits significant advantages over previous advanced methods in Ship Detection of varying scales in dense scenes. The code will be made available on GitHub.

Fine-Grained Visual Categorization (FGVC) aims to distinguish between extremely similar subordinate-level categories within the same basic-level category. Existing research has proven the great importance of the discriminative features in FGVC but ignored the contributions for correct classification from other features, and the extracted features always contain more information about the obvious regions but less about subtle regions. In this paper, firstly, a novel module named forcing module is proposed to force the network to extract more diverse features for FGVC, which generates a suppression mask based on the class activation maps to suppress the most distinguishable regions, so as to force the network to extract other secondary distinguishable features as the final features. The forcing module consists of the original branch and the forcing branch. The original branch focuses on the primary discriminative regions while the forcing branch focuses on secondary discriminative regions. Secondly, in order to solve the problem that information of small-scale distinguishable features is lost seriously after multi-layer down-sampling, according to the class activation maps of the first prediction, the object is cropped and scaled as the second input. To reduce the prediction error, the first and second prediction probabilities are fused as the final prediction result. Experimental results indicate that the proposed method not only outperforms the baseline model by a large margin (3.7%, 5.9%, 3.1% respectively) on CUB-200-2011, Stanford-Cars, and FGVC-Aircraft, but also achieves state-of-the-art performance on FGVC-Aircraft.