Leaf Recognition Using Deep Convolutional Neural Network

Abstract

Investigating its surroundings is a fascination for mankind that stems from human curiosity. By an estimate on our planet Earth, out of 8.7 million species there a total of 400,000 plants species and out of these at least 12,000 are economically valuable. Since the beginning of modern human society correctly identifying such a large number of plants has been a daunting task. In modern botanical sciences identification of plant is done using traditional text-based taxonomic keys based on leaf characteristics. It involves navigating the key containing following a series of decision making steps for identification of plants, but this is a tedious process requiring specialized domain knowledge primarily relying on field experts and scientists. In view of the several threats to general irreversible decline of plant biodiversity, speeding the task of identification of plants and making it accessible for non-experts would be highly beneficial. After recent developments in field of image processing an autonomous system for plant identification is in high demand. Experts in image processing have used leaf as a reasonable object to classify plants. Many techniques including Speed-Up Robust Features (SURF), Scale Invariant Features Transform (SIFT) and Histogram Oriented Gradient (HOG) have been used to extract features from leaves which include ratio, circularity, eccentricity, roundness, vein structure, color, texture and shape of a leaf centroid and contour distance. The obtained numerical values of these features are than provided to some form of learning algorithm like ANN or SVM for classification. The issue is that these methods are too specific because performance depend heavily on underlying hand engineered features. Deep neural networks, the current state of the art in A.I, especially deep convolutional neural networks can be used to solve this problem. These bio-inspired neural networks work on the same mechanism as human brain’s visual cortex and are capable of extracting features suitable for classification by themselves. Such a CNN model called Inception developed by Google, which was trained to classify 1000 categories of ImageNet dataset and was the winner of ILSVRC 2014 challenge was used for leaf classification through transfer learning. Modern object recognition models like Inception have millions of parameters and can take weeks to fully train. Transfer learning is a technique that shortcuts a lot of this work by taking a fully-trained model and retraining it on a totally new dataset under the notion that new datasets can be learnt easily by preserving and reusing the features learnt by the model from previous datasets. Therefore, the last fully connected layer of the Inception model, the Softmax layer, was retrained on four different leaf datasets separately. Three datasets used were FLAVIA dataset, Swedish leaf dataset and Leaf Snap dataset. The fourth dataset consisted of leaves of local Pakistani plant varieties. the results obtained after training and evaluating all four datasets were remarkably good and easily surpassed the previous best results achieved on these datasets.