Corn Diseased Leaf Segmentation Under Different Illumination Conditions

In order to study the influence of shadow on corn disease segmentation in a real environment, reduce the influence of shadow region on segmentation and determine the method suitable for segmenting corn disease in this environment, this study compares and analyzes the segmentation performance of traditional and deep learning semantic segmentation algorithms on the three datasets of SDD, NLDD and SNLHDD. In the first step, the image of corn leaf disease was collected using an RGB camera, which was divided into three categories according to the combination type of shadow area and normal illumination area. As a second step, all the manually extracted single feature segmentation methods (OTSU algorithm and Excess green color threshold algorithm), manually extracted multi-feature fusion segmentation methods (K-means clustering algorithm and Random forest algorithm) and deep learning-based semantic segmentation methods, were employed in the segmentation tasks. To conclude, the constructed datasets were used to test the traditional semantic segmentation methods and to train, verify and test the deep learning model. Among all the methods used in the experiments, the DeepLabV3+ deep learning model with ResNet18 as the feature extraction method provided the best results. With this method, MPA (SDD: 99.3%, HLDD: 99.1%, SHHDD: 97.9%) and MIOU (SDD: 78.9%, HLDD: 75.2%, SHHDD: 68.9%) were segmented very accurately. The model size was 58 MB, and the segmentation speed was 4 ms per image, which can satisfy the actual segmentation requirements. In addition, it is found that the model has a good segmentation effect on a single normal light or shadow, but the segmentation performance decreases when the normal light and shadow are mixed. It is proved that dividing the shadow area in the image separately and then segmenting the disease can reduce the influence of the shadow area on the whole image segmentation.