Plant leaves can be used to effectively detect plant diseases. However, the number of images of unhealthy leaves collected from various plants is usually unbalanced. It is difficult to detect diseases using such an unbalanced dataset. We used DoubleGAN (a double generative adversarial network) to generate images of unhealthy plant leaves to balance such datasets. We proposed using DoubleGAN to generate high-resolution images of unhealthy leaves using fewer samples. DoubleGAN is divided into two stages. In stage 1, we used healthy leaves and unhealthy leaves as inputs. First, the healthy leaf images were used as inputs for the WGAN (Wasserstein generative adversarial network) to obtain the pretrained model. Then, unhealthy leaves were used for the pretrained model to generate 64*64 pixel images of unhealthy leaves. In stage 2, a superresolution generative adversarial network (SRGAN) was used to obtain corresponding 256*256 pixel images to expand the unbalanced dataset. Finally, compared with images generated by DCGAN (Deep convolution generative adversarial network). The dataset expanded with DoubleGAN, the generated images are clearer than DCGAN, and the accuracy of plant species and disease recognition reached 99.80% and 99.53%, respectively. The recognition results are better than those from the original dataset.

The condition of plant root systems plays an important role in plant growth and development. The Minirhizotron method is an important tool to detect the dynamic growth and development of plant root systems. Currently, most researchers use manual methods or software to segment the root system for analysis and study. This method is time-consuming and requires a high level of operation. The complex background and variable environment in soils make traditional automated root system segmentation methods difficult to implement. Inspired by deep learning in medical imaging, which is used to segment pathological regions to help determine diseases, we propose a deep learning method for the root segmentation task. U-Net is chosen as the basis, and the encoder layer is replaced by the ResNet Block, which can reduce the training volume of the model and improve the feature utilization capability; the PSA module is added to the up-sampling part of U-Net to improve the segmentation accuracy of the object through multi-scale features and attention fusion; a new loss function is used to avoid the extreme imbalance and data imbalance problems of backgrounds such as root system and soil. After experimental comparison and analysis, the improved network demonstrates better performance. In the test set of the peanut root segmentation task, a pixel accuracy of 0.9917 and Intersection Over Union of 0.9548 were achieved, with an F1-score of 95.10. Finally, we used the Transfer Learning approach to conduct segmentation experiments on the corn in situ root system dataset. The experiments show that the improved network has a good learning effect and transferability.

For trees, leaves are often used for identification, but the shape of leaves changes greatly, bark will be another identifying feature. However, it is difficult to recognize by a single organ when there are intra class differences and inter class similarities between leaves or bark. So we fuse features of leaf and bark. Firstly, we collected 17 species of leaves and bark of trees through field shooting and web crawling. Then propose a method of combining convolution neural network (CNN) with cascade fusion, additive fusion algorithm, bilinear fusion and score level fusion. Finally, the features extracted from the leaves and bark are fused in the ReLu layer and Fully connected layer. The method was compared with single organ recognition, Support Vector Machines (SVM), and existing fusion methods, results show that the two organ fusion method proposed are better than the other recognition methods, and recognition accuracy is 87.86%. For similar trees, when it is impossible to accurately determine its species by a single organ, the fusion of two organs can effectively improve this situation.

To self-calibrate the intrinsic camera parameters, we designed a calibration pattern and proposed a method to estimate camera parameters involving distortion coefficients from orthogonal vanishing points. We first obtained four pairs of vanishing points from an image of the calibration pattern and selected the two pairs with the smallest error. Then, after analyzing the effect of the intrinsic parameter error from the noise of the vanishing points, we employed an optimization algorithm to obtain the best vanishing points. Using two images with different orientations, we could obtain the intrinsic parameters by using a linear method. Finally, according to the preliminary calibrated intrinsic parameters and a criterion that the airline is a static straight line in the pictorial plane, the arithmetic mean of the camera distortion, which is based on the common characteristic points, could be obtained. Simulated and experimental results indicated that the two-dimensional reconstruction error is 0.32 pixels. Although this method has the same error level as the traditional method, it has a better flexibility and can achieve real-time camera self-calibration.