Image recognition of Camellia oleifera diseases based on convolutional neural network & transfer learning

Abstract: Leaf diseases are a serious problem in Camellia oleifera production. The occurrence of Camellia oleifera disease is affected by various factors, such as variety, cultivation environment, climate condition and management level. The key to effective prevention and cure of Camellia oleifera disease is to identify the disease type timely and accurately. Traditional computer vision methods for plant leaf disease recognition depend heavily on time-consuming and elaborate feature design. To solve this problem, a recognition model of Camellia olerfera leaf diseases based on convolutional neural network was proposed and transfer learning was used to improve model's performance. Deep convolutional neural network has powerful capacities of feature learning and feature expression, which was used to learn features of diseased Camellia oleifera leaves. Transfer learning method was used to transfer the knowledge learned from ImageNet dataset by AlexNet to the identification task of Camellia oleifera diseases. The proposed model was implemented with Python programming language under the deep learning framework of Tensorflow by modifying the output number of the last fully connected layers in AlexNet to 5. We collected Camellia oleifera leaves in artificial Camellia oleifera land and took photos by mobile phone in bright indoor environment after flattening leaves. Leaf images were first converted from RGB (red, green, blue) color space to HSI (hue, saturation, intensity) color space, and then background was removed by threshold segmentation on hue and saturation channels. After segmentation, morphological open and close operations with a radius of 3 pixels were performed to remove burrs, holes and other noises, and thus the leaf mask was obtained by filling holes. Leaf mask was multiplied with the original image to obtain the colored leaf region. The colored leaf region was then rotated according to its principal axis angle and aligned horizontally. Based on the long edge, leaf image was scaled to 256×256 pixels. After these pretreatments, Camellia oleifera leaf images were manually identified as algal spot, soft rot, sooty mould, yellows and healthy leaf. A total of 750 images for each disease category were selected to form data set, 80% of samples were randomly selected for train set, and the remaining 20% for test set. To simulate different views of image acquisition and reduce over-fitting of network models, image datasets of diseased Camellia oleifera leaf were augmented by random crop, random rotation and random perspective transformation. To save space for huge amount of augmented images, data augmentation was executed online when training. In random crop mode, image is randomly cropped from 256×256 to 227×227 pixels. In random rotation mode, image is randomly rotated by 0, 90, 180, or 270 degrees. In order to avoid serious distortion of the transformed image, the displacement of the corresponding point in perspective transformation is limited to 10% of the image width and height. A total of 54 experiments were performed on Nvidia GPU with a combination of 2 learning methods (training from scratch, transfer learning), 3 data augmentation modes (no augmentation, random cropping, sequential execution of random cropping, perspective transformation and rotation), 3 regularization coefficients (0.0, 0.0005, 0.0001), and 3 initial learning rates (0.001, 0.005, 0.01). When training from scratch, weights are randomly initialized with truncated normal distribution and biases are initialized with zero constant. In transfer learning, only the last fully connected layers' weights and biases are reinitialized with random values, and those of other layers are assigned by the values from pre-trained AlexNet model. Experimental results show that transfer learning can significantly improve models' convergence speed and classification performance, and data augmentation can enrich data diversity and avoid over fitting especially when training from scratch. The classification accuracy was as high as 96.53% in transfer learning, and the F1 scores of algal spot, soft rot, sooty mould, yellows and healthy leaf achieved 94.28%, 94.67%, 97.31%, 98.34% and 98.03% respectively. This method has high recognition accuracy, and strong robustness to translation and rotation, and can provide references for intelligent diagnosis of plant leaf diseases.