ML Image Identifier Lite 1.3.0

FEATURES:

• ML Image Identifier is an educational application that enables real-time image identification on iOS devices as they are moved around the environment. It can recognize images in three main categories: "Objects," "Cars," and "Food," as well as identify "Text" content (character boxes, OCR), and recognize "Faces" by detecting feature landmarks.
• The app adjusts image processing levels automatically to function on any iOS 12 device, although performance may be impacted on older models. Devices with iOS 13 benefit from optical character recognition (OCR) in the "Text" mode.
• For categorized images, the application presents the top five predicted matches based on confidence levels derived from neural networks as percentages.

BACKGROUND:

Machine learning, once a concept confined to science fiction, has become an integral part of our daily lives over recent years. Its applications range from handwriting and facial recognition to image tagging, AI implementations in games, targeted advertising, predictive typing, and various automated processes. The value obtained from user-generated data in platforms like social networks underscores the notion that information equates to power.

With the introduction of iOS 11, Apple democratized machine learning through CoreML, facilitating the execution of neural networks and other ML tools via hardware acceleration on all iOS devices.

The ML Image Identifier serves as a demonstration showcasing the potentials and limitations of machine learning technology. Building an efficient neural network represents only a fraction of the overall task. Successful model operation requires substantial volumes of diverse test data analogous to the learning process experienced by living beings. The quality of test data plays a pivotal role: superior data results in superior outcomes, while inferior data might lead to subpar results. Furthermore, biases of those crafting the tests can inadvertently influence the outcome by giving precedence to specific test values over others.

SPECIFICS:

• "MobileNet" - Dedicated to scanning common objects, this model excels in identifying household items but lacks the ability to recognize people. Noteworthy for its high-quality inference capabilities in comparison to other voluminous ML models that can reach sizes of up to 500MB.
• "CarRecognition" - Designed for identifying vehicle makes and models, this model exhibits a hit-or-miss performance, with tendencies to match automobiles predominantly from certain global regions. Accuracy issues are prevalent, typically detecting the correct body type but misidentifying the make.
• "Food101" - Specialized in recognizing prepared foods, this model struggles with general food items and predominantly focuses on culinary offerings that might be uncommon at most households, like caviar and lobster. Known for frequent false positives when assessing desserts.

The text-recognition mode scans for all textual content within view and conveniently highlights words and individual characters for enhanced readability. Support for OCR functionality is extended on iOS13.

The facial-recognition mode identifies human or human-like faces within view and accentuates facial landmarks such as eyes, nose, and jawline. This mode performs optimally on newer devices due to enhanced hardware capabilities required for real-time image processing.