A typical hyperspectral imaging food inspection system contains a hyperspectral camera (typically on near infrared spectral range) and a machine learning software.
The hardware is mounted above a conveyor belt in food processing facilities, in order to capture images of products in real-time, as they pass below. The software models then analyze in real time those images to provide insights about the quality of the product instantly and non-invasively.
The adoption of hyperspectral imaging for on-line quality control and food safety applications requires advanced science and engineering in the following areas:
1. Processing Speed
- The speed of the algorithm is essential for on-line system deployment. Food companies continuously try to increase the speed of their conveyor belts to optimize throughput, which requires an inspection system where analysis of a hyperspectral image (which can be up to 1 GB in size) needs to occur in real time.
- We can apply advanced graphics processing units (GPUs) and machine learning techniques to increase the speed of their algorithm by dividing the image into subparts so classification of the whole image can happen in parallel.
- The big data volume of HSI has been a huge challenge for real time on-line food inspection, and that is mainly why although hyperspectral imaging technology has existed for over 20 years but could not be applied in food processing facilities for in-line use until recently.
- How to achieve an efficient algorithm that can be both fast and accurate? Leveraging GPU computation, designing compact fully convolutional neural networks that are capable of processing large HSI in less than one second is required.
2. Classification Data
- At Opsyne, we focus on optimizing spectral resolution and signal-to-noise ratio to identify and distinguish the most nuanced chemical attributes. For each pixel we can generate a spectral response that is in principle unique to each material or molecule. In other words, each element or feature absorbs light differently. With our hardware, we can identify these differences but they must be correlated to the known chemical attributes. These correlation datasets are known as classification libraries or ground truth data.
- To build the correlation between the spectrum of food products and the concentration of food ingredients, e.g., moisture or fat content, some ground truth food samples with lab-tested spectral and ingredient concentration values are required to train the machine learning models.
- Different food companies and food products may have specified ingredients that need to be quantified and require different sets of training samples. This, in turn, requires product-specific knowledge about how to operate the ground-truth reference measurement, and how this correlates with spectral data from a hyperspectral image.
- How to reduce the cost and difficulty of collecting abundant training samples, by designing advanced machine learning algorithms that can efficiently capture signature information from limited training samples and generate efficiently the learnt information to data of other companies or other food products? Advanced weakly supervised learning and domain adaptation approaches can be designed to achieve this purpose by fully exploiting the information and knowledge in the existing dataset via efficient learning and data augmentation.
3. Spectral Resolution
- The discriminative features that are sensitive to food ingredient concentrations tend to be hiding in high-dimensional hyperspectral spectral space, and thereby the extraction of discriminative features can be better achieved by exploiting the fully spectral information using all spectral channels as input.
- The current approaches tend to use limited “importance” channels for food ingredient quantification. For example, the band centered at 1435nm is used for crystalline sucrose quantification because crystalline absorbance peaks at 1435nm. Similarly, bands centered on 1724 and 1762 nm (CH2 absorbance) are used for fat content quantification, and several water absorbance bands (including at 1925 nm) are used for moisture quantification. So, in current approaches, many channels are ignored and wasted, and the advantage of rich spectral channels in hyperspectral imaging is not leveraged.
- Moreover, the knowledge for “important” channel selection is not universal and will be biased if the circumstance changes with product types, ingredient types, and camera types. So, these knowledge-driven “important” channels selection approaches are not adaptive to different food products, different ingredients and different cameras. With the state-of-the-art optics offered by the Opsyne hardware, these challenges can be overcome.
- How to use full spectral information, such that more discriminative features can be extracted for enhanced food ingredient quantification and mapping? The deep neural network approach that ingests all spectral channels and automatically generates the most discriminative features in a data-driven manner will be helpful.
Please contact us at firstname.lastname@example.org to learn how advancements in robotics and machine learning combined with our advanced optical engineering is significantly improving the economics and efficacy of automated hyperspectral imaging solutions across the food industry and throughout the manufacturing process.