by Melissa Kimbrell, Trainer and Technical Support Specialist at Amped Software
Artificial Intelligence (AI) has become a hotly debated topic in many different industries. It has exploded in popularity over the last few years, largely due to increased computing power and overall technological improvements. AI is being leveraged for all kinds of tasks, some of which we are largely familiar with, such as writing and illustrating, online shopping, vehicle navigation, and social media. Others might be less obvious, such as computer coding, diagnosing cancer, learning a new language, video gaming, curating a playlist of music you’ll love, and robot vacuums navigating your home.
AI in Legal and Forensic Applications
When it comes to law enforcement and legal proceedings, there are mixed opinions. In the landmark case of State of Washington v. Puloka, the court rejected the use of AI-enhanced imagery, ruling it inadmissible because it failed to meet the admissibility standard of general acceptance in the relevant scientific community. Amped Software agrees with this ruling and maintains a firm stance that AI does not currently have a place in the restoration and enhancement of image-based evidence. Evidence processing, like any forensic science discipline, requires reliability, repeatability, and reproducibility. Since one cannot know for certain how a machine-learning algorithm has been trained, it is impossible to know reliably the process by which it determines how to clarify an image and improve it. Each time this operation is performed, it could potentially result in a different solution, thereby failing a test of repeatability. Consequently, because the method is neither reliable nor repeatable, another qualified analyst could never expect to reproduce the same result exactly.
Amped Software believes AI has its successful place in the investigative stage. At this point, any leads generated will be fully scrutinized before anything is presented in a court of law. In the case of Facial Recognition, for example, a series of possible facial matches might be returned to an investigator. Each lead must, however, be researched and interrogated independently. While the result of the Facial Recognition search would not be admitted as evidence, it does provide a vital step in furthering an investigation.
When video footage relevant to an accident or crime scene is evaluated, oftentimes license plates become visible. While investigators approach these scenarios with optimism that the license plate characters will be clearly legible, this is not often the case. Though cameras and recording systems are increasing in quality and resolution, video examiners are continually facing heavily compressed or poor-resolution license plates that are difficult to read. This is where Amped Software’s DeepPlate can shine.
Introducing Amped Software’s DeepPlate
DeepPlate is a deep-learning-based algorithm for deciphering license plates affected by the common issues introduced by surveillance systems: perspective distortion, poor resolution, optical and motion blurring, and compression noise. Amped Software trained its dedicated neural network with millions of synthetically generated and distorted license plates utilizing the known font, spacing, and character structure configurations for several countries and states. Since the source plates were synthetically generated, Amped does not have, nor use, any stored data related to any real license plates.
While Amped Software wanted its Amped FIVE users to have access to this investigative tool, it was adamant about keeping all AI separate from the mathematical basis of Amped FIVE. This separation aligns with Amped’s belief that AI does not currently have a place in the restoration and enhancement of imagery evidence. Therefore, Amped provides 50 uses per month (per license seat) within the Support Portal to users with an active license of FIVE.
How to Use DeepPlate
The usage of the tool is easy!
- Navigate to the Support Portal and select the DeepPlate tab.
- Agree to the Terms and Conditions to access the Upload page.
- Choose an available license from the dropdown menu, ensuring the selected license has remaining DeepPlate uses for the month.
- Select the relevant country. Some countries offer additional dropdowns for narrowing down to specific states or territories. Selecting a state requires the specific configuration of the letters and numbers in the license plate in question. Since these can often be customized or non-standard, the State selection can be omitted.
As an example, this image displays an enhanced frame from a surveillance video:
Because it is suspected to be a standard Texas license plate, the following selections were made:
- Select the image in the “Choose File” dialog box and proceed by hitting the “Upload” button. Note that though DeepPlate does require the uploading of the file, it will not be stored beyond the use of the DeepPlate process.
- Once the image is uploaded, select the four corners of the license plate by right-clicking each corner, beginning with the top left corner and proceeding in a clockwise manner. If needed, press the “Clear selection” button to perform the right-click selection again.
- Press “Continue” for DeepPlate to run its process.
- The tool rectifies the image by enlarging and correcting the perspective of the license plate. Before viewing the results, users are reminded that this is an investigative tool only and should not be relied upon as evidence. It is encouraged to form independent conclusions prior to viewing the results for bias mitigation.
- Click “Show results” to view two charts.
- The first chart lists possible characters for each position, sorted by confidence level. The confidence level is only how confident the neural network is about its conclusion. A high confidence level does not mean you can be sure the character is correct. As is the case with any neural network, DeepPlate can be very confident about a character and still be incorrect.
- The second chart is derived from the first one. It displays a list of 60 possible license plates sorted by the aggregated confidence of the characters. This is computed by multiplying together the individual confidence score of each character in the license plate.
- By the time you see the results, the data deletion process has already begun on Amped Software’s servers. The results page is stored locally on your browser cache, but at this point, the imagery is no longer retained. The “Generate PDF” button exports the results to a PDF file for later use. Results will be presented on the second page of the PDF to mitigate bias as much as possible.
Final Considerations
Amped Software has been very pleased with the ability to provide this tool to its users around the world. Development is ongoing to expand support for more countries and territories in the future. Amped Software always prioritizes clarity about its capabilities and intentions. For this reason, it is transparent about this process being AI-based, and maintains that this tool cannot replace a human in making conclusions regarding license plate characters, even only in an investigative context. It should also be noted that there are times when DeepPlate will be inexplicably incorrect, which is exactly the lack of explainability that leads Amped Software to exclude the use of deep-learning algorithms from its evidentiary material.
