[month] [year]

Vamsidhar M – Computer Vision for Flora

December 2022

Vamsidhar Muthireddy received his Master of Science in  Computer Science and Engineering (CSE).  His research work was supervised by Prof. C V Jawahar. Here’s a summary of his research work on Computer Vision for Flora: Species and Disease Recognition.

Identifying plant species by the visual characteristics of leaves, flowers, and seeds is a crucial key component in the conservation of endangered plants. These traditional identification methodologies are manual, time-consuming, and require domain knowledge to operate. With the growth of technology, the necessity for automated plant identification systems is evident. When integrated with mobile platforms like smartphones, such a system can aid plant-related education, promote ecotourism, and create a digital heritage for plant species, among many others. For plant species classification, we utilize the convolutional neural networks. Image datasets required for classification are not easily available for plant species. We take this opportunity to compose the Indic-leaf dataset. This dataset contains images belonging to 112 Indian plant species, wherein each plant species is captured in different resolutions. We explain the reasoning behind it and use the resolution information of the image to gain insights into the type of data that affects the plant species classification in the wild. The P@1 score obtained for the classification task on our dataset is 90.08. The best-performing model is then utilized on other existing plant species datasets to show the complex nature of the Indic-Leaf dataset. We then discuss and elaborate on our crowd-sourcing web application to collect and regulate the data. We explain how the automated plant identification system can be integrated with a smartphone by detailing the flow of our mobile application. Diseases in the plant species often cause changes to their visual characteristics. These distinct morphological features assist in distinguishing a diseased leaf from a healthy one. Most of the diseases do not correspond to a specific plant species. When such diseases manifest themselves across different plant species, they often have similar visual characteristics. However, having prior knowledge of the plant species would aid in accurately identifying the disease. We conduct our analysis of the diseases by classifying plant-disease pairs. For this, we use an existing dataset that is similarly structured. We propose a hybrid model to utilize the hand-crafted features in our classification process. This hybrid model makes use of data fusion techniques to inject these features into the convolutional neural network. Experimentations demonstrate the efficacy of our proposed model for plant disease classification. We obtained the P@1 score of 99.74 using our proposed hybrid model against the baseline P@1 of 99.34. Visualization methodologies are used to obtain the qualitative results for the classification models. From the obtained results, we observe that a significant amount of the feature space for the classification task is obtained from the background objects. This trend is observed in deeper ResNet architectures when trained on scan-type images. The quantitative analysis performed back our observation. We note that a neural network should also be assessed based on its selection of the feature space. A neural network should utilize more features extracted from the foreground than the background while performing the classification task. Necessary modifications are proposed to incorporate this into the assessing criteria of the classification models. We evaluate the various classification algorithms that we utilized using the proposed methodology. The results show the necessity of assessing the convolutional neural networks based on their feature space.