Quantifying plant infestation by estimating the number of biological objects on leaves, by convolutional neural networks that use training images obtained by a semi-supervised approach
A computer generates a training set with annotated images ( 473 ) to train a convolutional neural network (CNN). The computer receives leaf-images showing leaves and biological objects such as insects, in a first color-coding ( 413 -A), changes the color-coding of the pixels to a second color-coding and thereby enhances the contrast ( 413 -C), assigns pixels in the second color-coding to binary values ( 413 -D), differentiates areas with contiguous pixels in the first binary value into non-insect areas and insect areas by an area size criterion ( 413 -E), identifies pixel-coordinates of the insect areas with rectangular tile-areas ( 413 -F), and annotates the leaf-images in the first color-coding by assigning the pixel-coordinates to corresponding tile-areas. The annotated image is then used to train the CNN for quantifying plant infestation by estimating the number of biological object such as insects on the leaves of plants.
1 . A computer-implemented method for generating a training set with annotated images, wherein the annotated images are to be used to train a convolutional neural network (CNN) for quantifying plant infestation by estimating a number of biological objects on leaves of plants, the method comprising:
receiving a plurality of leaf images showing the leaves and showing the biological objects on the leaves, the leaf images being coded in a first color-coding;
changing the color-coding of the pixels in the leaf images to a second color-coding, the pixels in the leaf images including pixels for the biological objects and pixels for the leaves, wherein a contrast between the pixels for the biological objects and the pixels for the leaves is higher in the second color-coding than in the first color-coding;
assigning the pixels in the leaf images associated with a first color cluster in the second color-coding to a first binary value, and the pixels in the leaf images associated with a second color cluster to a second binary value;
differentiating the leaf images having binary coding, comprising areas with contiguous pixels in the first binary value into non-object areas and object areas by an area size criterion;
identifying pixel coordinates of the object areas of the pixels of the leaf images, wherein the pixel coordinates identify rectangular tile areas with biological objects center centered therein, with the tile areas having an input dimension for the CNN; and
annotating the leaf mages in the first color-coding by assigning the pixel coordinates to the corresponding rectangular tile areas to obtain the annotated images.
2 . The method of claim 1 , wherein receiving the leaf images comprises receiving images showing isolated segmented leaves, wherein each leaf image of the leaf images shows the leaf with its margin substantially completely.
3 . The method of claim 2 , wherein receiving comprises obtaining the images with the isolated segmented leaves by performing leaf segmentation using the CNN that has been trained by processing leaf-annotated plant images that show annotated borders between leaf and background, and between leaf and leaf.
4 . The method of claim 1 , wherein the first color-coding is RGB, and wherein changing the color-coding comprises a transformation from the RGB to XYZ.
5 . The method of claim 4 , wherein changing the color-coding further comprises obtaining the second color-coding by disregarding a Z-component from the XYZ color space.
6 . The method of claim 5 , wherein assigning is performed by clustering the pixels in the leaf images into color clusters.
7 . The method of claim 6 , wherein the color clusters are identified by a support-vector machine.
8 . The method of claim 6 , wherein identifying the pixel coordinates is performed with the rectangles of the rectangular tile areas being squares of 96×96 pixels.
9 . The method of claim 1 , further comprising classifying the rectangular tile areas according to object classes and sorting out false positives.
10 . The method of claim 1 , wherein
the biological objects are insects,
the non-object areas are non-insect areas, and
the object areas are insect areas.
11 . The method of claim 1 , wherein relation of physical size of the biological objects to the physical size of the leaves is such that a representation of the biological objects on the leaf images are such that the representation is smaller than the dimension of the tile areas.
12 . The method according of claim 11 , wherein the biological objects are selected from the following: insects, arachnids, and mollusca, or wherein the biological objects are selected from the following: spots or stipes on a surface of plant parts.
13 . A computer adapted to perform the method of claim 1 .
14 . A non-transitory computer-readable medium having a computer program product encoded thereon that, when loaded into a memory of a computer and executed by at least one processor of the computer, causes the at least one processor to perform the method of claim 1 .
15 . A computer-implemented method for quantifying plant infestation by estimating a number of biological objects on leaves of a particular plant, the method comprising:
receiving a plant image taken from the particular plant, the plant image showing at least one of the leaves of the particular plant;
using a first convolutional neural network to process the plant image to derive a leaf image comprising a contiguous set of pixels that show a main leaf of the particular plant completely, the first convolutional neural network having been trained by a plurality of leaf-annotated plant images, wherein the plant images are annotated to identify main leaves;
splitting the leaf image into a plurality of tiles, the tiles being segments of the plant image having pre-defined tile dimensions;
using a second convolutional neural network to separately process the plurality of tiles to obtain a plurality of density maps having map dimensions that correspond to the tile dimensions, the second convolutional neural network having been trained by processing object-annotated plant images, wherein a first subset of the object-annotated plant images is obtained by interacting with an expert user, and a second subset of the object-annotated plant images is obtained by performing the method of claim 1 , the processing comprising the calculation of convolutions for each pixel of the contiguous set of pixels based on a kernel function resulting in density maps with different integral values for tiles showing the biological objects and tiles not showing the biological objects; and
combining the plurality of density maps to form a combined density map in the dimension of the leaf image, and integrating the pixel values of the combined density map to obtain an estimated number of the biological objects for the main leaf.
16 . The method of claim 15 , wherein the biological objects are insects, and the object-annotated plant images are insect-annotated plant images.