LeavesDataset

INTRODUCTION

Automated leaf segmentation is basic for many agriculture-related projects like disease detection from plant leaf, leaf classification or identification etc. But at the same time, it is an extremely difficult part. Due to variability in colour, shape and size of leaves, it becomes extremely difficult to have a general method for all species of plants. So here we have focused on only one species. The aim of this report is to get leaves from an image as precise as possible rather than getting all leaves present in the image. Special care has been taken to identify background as precisely as possible. Here part of the image containing plant is known as foreground while the remaining part is known as background. In this report, we have discussed leaf segmentation for specific plant i.e. brinjal plant.

DATA

We have used images of eggplants captured from 13MP smartphone camera with 1.34-micron pixel size and an f/2.2 aperture. All the images show the top view of plants with 3120 x 4160 pixels resolution. Images vary from young to old plants within a duration of two months. Images are also taken during a different time of day. Some images are taken in the morning, some during noon and some in the evening. Some images are taken before watering of plant and some after watering. In this data, we have tried to cover all different combination possible which effect size, shape and colour of the plant. Total of 100 images of eggplants is used to extract leaves from each image. Each eggplant image contains on an average 15 leaves.

METHODOLOGY AND APPROACH

Here first two steps extract the foreground from an image. As shown in figure 1 in the first step we are forming a cluster of pixels which have the same characteristic using superpixel clustering [1]. For superpixel segmentation, we are using simple linear iterative clustering (SLIC) algorithm.

As eggplant are green in colour, we are extracting superpix- els with green mean colour from the image. So these extracted

superpixels collectively form foreground. If the background is grassy surface then it can be wrongly identified as a foreground but this problem will be taken care in the fourth step. After extracting foreground edge is detected by using canny edge detection so we can separate each leaf. Now to find seed for each leaf we select such pixel which is present in foreground and it is not an edge. After finding seed by simple region growing approach mask is formed. If the mask contains some hole then it is corrected by filling that hole. This can be due to some disease or dryness of leaf which is not green in colour. If the size of the mask is greater than a particular threshold then all the superpixels which cover this mask are selected as leaf image. This threshold is useful to identify the difference between grass and leaves. As the mask of grass will be much smaller in size than that of leaves.

[[Add Image Here]]

Simple linear iterative clustering (SLIC)

Superpixels are a group of pixels with the same character- istics. Superpixel segmentation is local clustering of pixels in

5D space (labxy). Where xy is the position of pixels while the lab is a vector in LAB colour space. In LAB colour space l component represent lightness of pixel from 0 to 100. a component represents green to red colour while b component represents blue to yellow. a and b component value varies from -128 to 128. As we are taking lab value in consideration it can easily differentiate between foreground and background. It also differentiates between different leaves due to high lab value of edges. A number of segments is the only parameter in the SLIC algorithm. Now suppose if there are N pixels in a given image and we want M superpixels. As superpixels are almost equal in size due to its confined search area. So each segment will have approximately N

M pixels. Hence each cluster center will

be R pixels apart as shown in figure 2. R = r N M

(1)

Here N represent number of pixels in image, M represent required number of superpixels As shown in figure 3 search area of local clustering for a particular centre is 4R2 pixels surrounding that centre. So pixels from only this search area can be assigned to that particular centre. In Euclidean distance, position distance can reduce the effect of lab vector distance. So SLIC algorithm instead of using normal distance measure it uses normalized distance measure.

D = s Dxy 2 Pmax 2 + Dlab 2 Lmax 2

(2) Here Dxy represent euclidean position distance between pixel and centre while Dlab represent euclidean lab vector distance between pixel and centre. Pmax represent maximum

distance between pixel and center and Lmax represent maxi- mum lab vector distance between center and pixel.

Now Pmax = R as maximum position distance in this search area is R and Lmax vary depending on the image so will replace it with constant L. The value of L mostly varies from zero to forty. In this report, we have used 25 for a value of L. Value of L represents the effect of lab vector distance in forming superpixels. By increasing value of L can increase effect of lab vector distance and by decreasing value of L can increase effect of pixel position distance in forming superpixels.

D = s Dxy 2 R2 + Dlab 2 L2

(3) In SLIC algorithm first cluster centres are assigned to each cluster as shown in figure 2 then each pixel is assigned to its cluster depending on the value of D with respect to the cluster centre. If pixel fall in more than one search area of centres then centre which has a minimum value of D with that pixel is assigned to that pixel. After assigning each pixel to its centres, centres are shifted to the pixel which has an almost average value of the labxy vector. Here error is defined which is an average of the distance between previous centres and new centres. This process is iterated till error is less than the defined threshold. Figure 4, 5 and 6 shows the superpixels after 5, 7 and 10 iterations. We have observed for this data that for most of images iteration required is between 10 to 15. The time complexity of this algorithm is O(NI), where I represent a number of iteration, but as I is always less than 15-time complexity of this algorithm is O(N).

Extract foreground

Eggplants are green in colour, so to extract foreground all superpixels with green mean colour are extracted. In LAB colour space value less than -5 of a coordinate defines the green colour. All the segments with green mean colour are extracted. As discussed above background with grass can be identified as foreground. This problem is taken care of by extracting leaves with mask size more than a specific value. Some plant might have green colour value more than -5 which can be identified as background. But our main focus is to not identify background as foreground. Figure 7 shows extracted foreground for particular image.

Edge detection

After extracting foreground edge of leaves are detected using canny edge detection. Canny edge detection removes noise using the Gaussian filter. It is immune to a noisy environment. It is a well-defined method and offers reliable detection. Figure 8 shows detected edges on foreground figure

Leaf mask

Simple region growing approach is used to get a leaf mask. The seed for region growing is selected by simple iteration with condition that if the pixel is in the foreground and not an edge. This leaf mask is then corrected if there are any holes in it due to disease or dryness. Figure 9 shows a leaf mask before correction.

Leaf extraction

At last all superpixels which fall in the region of this leaf mask are extracted. These superpixels collectively form a leaf. Figure 10 shows extracted leaf of leaf mask shown in figure 9

RESULT AND CONCLUSION

Results of this method depend on different parameters like the number of segments, Gaussian filter, a component values for the green colour, size of mask allowed etc. A number of segments depend on the size of the image as well as the distance between the camera and plant. If a number of segments are more than required than most of the superpixels will be of no use as it will form a cluster which is squared in shape. If most of the clusters are squared then there will be no much difference between pixels and superpixels. If the number of segments is less than required than some pixels can fall in more than one leaves which will again hinder the purpose of superpixels. In this method, we have taken the value of a number of segments equal to 800 by considering the above observations. As for the value of a component and size of the mask, our only concern is background should not be identified as foreground it becomes easy to find the required value. Size of the mask can be useful to detect the difference between leaves and grass. In this method, if mask size is greater than 4 superpixels then it is considered to be a leaf. In this method, we have seen that leaves identification are more accurate in older plants than younger plants due to its texture. The texture becomes lighter as the plant grows older which help in detecting edges with less spurious edges or weak edges. As the plant grows older midrib and veins become less visible in this plant. Weak edges are formed due to veins and midrib. Figure 12 shows leaves extracted from figure 11. The image used in the example is of the younger plant. Leaves other than shown in figure 12 are either not detected due to the size of the mask or due to spurious edges generated by its veins. As our main focus was to not detect background as foreground which can be seen from figure 12 is easily satisfied. We have created a dataset of eggplants with more than 100 images and more than 1500 leaves in total. This dataset can be used in agriculture-related projects like disease detection, leaf classification, leaf identification etc.