Application-aware image compression for low cost and distributed plant phenotyping

Abstract

Plant phenotyping investigates how a plant's genome, interacting with the environment, affects the observable traits of a plant (phenome). It is becoming increasingly important in our quest towards efficient and sustainable agriculture. While sequencing the genome is becoming increasingly efficient, acquiring phenotype information has remained largely of low throughput, since high throughput solutions are costly and not widespread. A distributed approach could provide a low cost solution, offering high accuracy and throughput. A sensor of low computational power acquires time-lapse images of plants and sends them to an analysis system with higher computational and storage capacity (e.g., a service running on a cloud infrastructure). However, such system requires the transmission of imaging data from sensor to receiver, which necessitates their lossy compression to reduce bandwidth requirements. In this paper, we propose an application aware image compression approach where the sensor is aware of its context (i.e., imaging plants) and takes advantage of the feedback from the receiver to focus bitrate on regions of interest (ROI). We use JPEG 2000 with ROI coding, and thus remain standard compliant, and offer a solution that is low cost and has low computational requirements. We evaluate our solution in several images of Arabidopsis thaliana phenotyping experiments, and we show that both for traditional metrics (such as PSNR) and application aware metrics, the performance of the proposed solution provides a 70% reduction of bitrate for equivalent performance.

The significance of image compression in plant phenotyping applications

“…There are several scenarios where such efficiency may be necessary. For example, when images are acquired in a greenhouse facility or even in the field, and are then transmitted to a central location for archival and analysis (e.g., as in the framework proposed by (Minervini and Tsaftaris, 2013), or in the gigapixel time-lapse panoramic imaging system in (Brown et al, 2012)). Another example could be the recent developments towards affordable phenotyping 2 where users in developing countries or in rural remote areas acquire images using affordable and low computational power devices (e.g., mobile phones), and transmit them over wireless communication links (enabled in remote places by long-distance connectivity projects (Murillo et al, 2015) or emerging technologies such as the Brck 3 ) and the Internet to cloud services (e.g., the iPlant Collaborative (Goff et al, 2011)), where sophisticated analyses can take place, and results are sent back in response (Minervini and Tsaftaris, 2013;Puhl, 2013).…”

“…For example, when images are acquired in a greenhouse facility or even in the field, and are then transmitted to a central location for archival and analysis (e.g., as in the framework proposed by (Minervini and Tsaftaris, 2013), or in the gigapixel time-lapse panoramic imaging system in (Brown et al, 2012)). Another example could be the recent developments towards affordable phenotyping 2 where users in developing countries or in rural remote areas acquire images using affordable and low computational power devices (e.g., mobile phones), and transmit them over wireless communication links (enabled in remote places by long-distance connectivity projects (Murillo et al, 2015) or emerging technologies such as the Brck 3 ) and the Internet to cloud services (e.g., the iPlant Collaborative (Goff et al, 2011)), where sophisticated analyses can take place, and results are sent back in response (Minervini and Tsaftaris, 2013;Puhl, 2013). Both of these scenarios involve: a remote sensing device, which does not have the computational power to perform analysis; the use of a limited communication channel, which may not have the capacity to carry many large images; and potentially imaging of plants in non-ideal settings, for example in the field (Andrade-Sanchez et al, 2013;Bucksch et al, 2014) or non-uniformly illuminated conditions, which increase the complexity of the image content.…”

“…However, these measures are not specific for plant experiments, where the ultimate information of interest is the actual measure describing a plant trait. Therefore, compression performance should be evaluated against how accurately the trait of interest can be measured (Minervini and Tsaftaris, 2013). As quantitative traits we exemplarily investigate image segmentation-based traits and traits based on image sequence analysis.…”

Image Analysis: The New Bottleneck in Plant Phenotyping [Applications Corner]

“…Then the transmission of (possibly) large volumes of image data necessitates compression to meet bandwidth constraints. While this loss of information will affect the accuracy of the analysis algorithm, recent advances in application-aware compression can tune compression parameters to meet analysis accuracy needs [13], [14]. From a software engineering perspective, backward compatibility of the analysis framework and of the computational backbone has to be ensured, such that experimental protocols and results obtained previously remain valid.…”

“…Our recent project [16] aims to provide a universal turnkey and modular platform based on a distributed sensing and analysis framework [13], as shown in Figure 2. This distributed approach presents several key advantages.…”