Checker: ImageJ plugin for the visual checking of True Positives, False Positives and False Negatives in AgroDeep studies

(c) 2023 P. Borianne / CIRAD
Version 2.0 - February 2023
________________________

Checker is based on ImageJ developed by Wayne Rasband and is distributed under CeCILL-B License.
________________________

Thanks to Wayne Rasband and everyone who wrote and shared ImageJ plugins.
________________________

User guide¶

Checker was developed in the framework of the AgroDeep (AD) project; it allows to visually analyze, and if necessary to correct, the prediction errors of a specialized neural network. It is distributed under a Cecill-B license.

Principle¶

The validation of a neural network consists, among other things, in comparing the expert's annotations to the network's predictions. This work leads to divide the objects into three distinct types:

• True Positives, made up of objects simultaneously annotated by the expert and detected by the network,
• False Positives, corresponding to objects detected by the network but not corresponding to any expert annotation,
• False negatives, corresponding to objects annotated by the expert but not detected by the network.

The result checking consists in making sure that the various objects present in the image are correctly "typed"; it is necessary to be able to correct errors that are not the responsibility of the network (for example FALSE FALSE POSITIVE which correspond to an expert annotation defect).
The Checker plugin allows you to quickly visualize the elements by type and if needed

1. to transfer an object from the False Positive type to the True Positive type, thereby correcting an expert annotation defect,
2. to delete an object of the type False Negatives, thereby correcting an excess annotation,
3. to reclassify an object, particularly in the case of multi-class networks.

Input / Output block¶

Here are specified the data model to be used and the folders containing the images and type files respectively.

• The Model file to use: this is a TXT file that specifies the name and color of the object classes present in the images.

• The folder Typed Annotations: it contains the TXT files of the typed annotations coming from the validation processes of the AgroDeep platform; the network validation evaluates the pairing between expert annotations and network predictions: a true positive (TP) is a network prediction matched to an expert annotation, a false positive (FP) is an unpaired prediction and a false negative (FN) is an unpaired annotation. False positives are either annotation defects of the expert or detection or classification confusions of the network; false negatives are are either detection defects of the network or annotation confusions of the expert.

• The folder Corrected Annotations: it contains the TXT files of annotations after taking into account the modifications made by the operator; these files can be used in the training processes of the AgroDeep platform.

• The folder Images: it contains Image files (png, jpg, JPEG,...) with the same radical as the files contained in the Annotations folder.

Check block¶

The filters used are specified here in terms of Type ("True Positive", "False Positive" or "False Negative") and Class. Figure 1 specifies that the check will focus on false positives of the "cow" class.

Target block¶

This section indicates what change will be explicitly made by the operator; Figure 1 indicates that the current "false positive" of the "cow" class will be transformed into the "true positive" of the "shadow" class.

Objects block¶

Here are proposed the functionalities applicable to the objects:

• < : access to the previous object of the current image in the specified type,
• > : access to the next object of the current image in the specified type,
• Change Type : allows to modify the type or class of the current object,
• Suppress : removes the current object from the list of objects associated with the current image,
• Undo : cancels the last operation performed.

Functionalities are blocked / unblocked depending on the filters used.

Images block¶

Allows to navigate through the images in the selected folder in the Input/Output block. The name of the current image is displayed, in blue if already seen, in red if never seen. Changing an image causes the changes made to the current image to be saved.

Displaying¶

Checker is a visual control tool: it displays the image being analyzed and the selected object in the form of a thumbnail, materialized in the image by a red rectangle. _A rectangle drawn in blue indicates that the object has already been seen. The control cycle is potentially complete for the selected category of the current image when no thumbnails are available or when the offered thumbnails are all displayed in blue..

Triggered operations (change of type or class, deletion) are applied on the thumbnail object.

Downloading / Installation¶

Prerequisite: Install the ImageJ or Fiji application on the computer.

Once downloaded, the Checker_.jar plugin must be copied to the plugins subfolder in the ImageJ location or installed using the GUI of ImageJ / Fiji. In both cases, the applications will have to be restarted.

Lauching¶

The Checker plugin is launched from the graphical interface of ImageJ or Fiji.

Citation¶

Borianne, P., 2020. http://amap-dev.cirad.fr/projects/checker/wiki

---