Images are an efficient way to convey information that people tend to trust as they are considered to be objective. Thus, images have become the most exchanged media on the internet. In the context of spreading fake news, it is easy to be fooled by misinformation carried by images. Furthermore, it is now effortless to create convincing fake or totally generated images without any technical background in graphical software programs or computer science. In this presentation, I will present image forensics bases and related topics, and we will discuss how to detect if an image is likely to convey misinformation, i.e., to detect if it is generated or if it has undergone manipulation. Detection approaches can be passive using no prior knowledge or active by adding a hidden mark in the image. The latter was mostly used in some specific cases, such as press photography, and was considered a mature technology. However, image generation has recently brought watermarking back into focus, especially since the regulations imposed by the USA and the EU. Also, we will show how image generation raises challenges related to machine learning such as domain adaptation, unbalanced prediction, and interpretability. Indeed, the variety of images makes the set of real images really hard to model, therefore, the majority of methods rely on training sets. Unfortunately, these sets are too specific and fail to apply in the wild in operational settings. Moreover, new models for image generation appear regularly and make these training sets obsolete. Finally, the interpretability of real or fake image classifiers is interesting to address for two reasons: it helps us to understand what biases they exploit and may increase explainability for the user.

Bioimage analysis is concerned with the quantitative analysis of biological image data. In this context images are often considered spatially distributed measurements. They can represent biological structures but also biological processes like translation, apoptosis and others. Imaging and automated image analysis are used as a tool in biological experiments. The scale of images of biological objects ranges from nanometers on the subcellular level to meters for individual animals and plants or even kilometers for whole ecosystems. However the inner workings of cells and the interactions between cells, forming tissues and organs, are a central part of biology. Individual cells are often on the micrometer scale and microscopes are used to make them visible. The field of epifluorescence microscopy is especially important for bioimage analysis. Fluorescent tags are attached to molecules of interest. The fluorophores are stimulated by a laser pulse with a given wavelength to emit photons, which are captured by the microscope. This way multiple molecules or structures can be imaged independently from each other. Optical microscopy is unable to resolve structures below a limit of approximately 200nm, due to the diffraction limit. Multiple approaches like SIM, STED and STORM exist to get around the diffraction limit. Life happens in a 3D world. With the help of optical slicing, 3D images can be taken. The confocal laser scanning microscope, the spinning disk microscope, the multiphoton microscope and the lightsheet microscope allow to capture 3D images with a high resolution. To obtain correct results, the bioimage analysis must in many cases take into account the specifics of the image formation process and the degradations introduced by the microscope used. Correcting these degradations is called image restoration. The image is considered as a superposition of the signal of interest, different forms of noise and background. For the epifluorescence microscope the point spread function (psf) is the impulse response of the optical system and the acquired image is a linear superposition of psfs. With the help of the psf, a less degraded image can then be approximated using image deconvolution. Other pre-processing operations include alignment or stiching, noise removal and background correction. In order to count biological objects and to measure their intensity and shape features, the objects must be detected or segmented. The calculation of features in the form of convolutional filters and ranking filters, thresholding and post processing of the resulting masks with the help of mathematical morphology are the basic steps of image analysis workflows. Machine and deep learning can provide better solutions. Pixel and object classifiers based on the random forest model are often used. When these do not succeed, convolutional neural networks can be trained to accomplish semantic or instance segmentation. Often networks based on the UNet architecture are used in bioimage analysis. The analysis can go beyond measuring the features of individual objects, for example by analyzing the spatial distribution of molecules within cells, the colocalization of different molecules or the movements of specimen, cells and molecules. We will demonstrate the application of bio-image analysis using two examples we recently worked on: The segmentation and detection of microglia on fluorescent images and the delineation of bark, pith, and xylem annual rings on stained sections of tree trunks.