Complexity bounded classification of fish-eye distorted objects with micro-controllers

Since 2019, tiny machine learning has imposed itself everywhere as an innovative technology trend deployed at the edge and has been pervasive in many IoT applications. One interesting, addressed by this work, is related to the welfare of the laboratory animals that could be preserved by acquiring and classifying some image data to monitor object’s presence in their cages. For example mice and rats activities such as drinking and eating can be indicators of their state of health. In that context, $\mu$BottleNet has been developed to classify the presence, or the absence of the water bottle, while $\mu$FoodNet to classify the level of the food into the feeder. Those neural networks (NNs) achieved 99.8% accuracy. At our best knowledge, no datasets were available to support this study and therefore we built three image datasets to train the neural networks on both Bottle and Food objects. From image capture to the inferences execution, the tasks have been carried out by STM32L4 (ultra-low-power that consumes 120 $\mu$A/MHz) and STM32H7 micro controller units (MCUs). Both NNs have been designed to fit into constrained MCU resources. Special attention has been given to the on-chip memory occupation to ensure the RAM footprint was 39.44 KBytes. To evaluate and test on the field the performances of these models (and against to the MobileNetV2 one), a graphical user interface (GUI), has been developed, capable of rendering, the validation and test results of inference runs on the MCUs. K-fold cross validation has been run and the resulting performances of the two NNs were compared to the MobileNetV2 confirming that the $\mu$BottleNet and the $\mu$FoodNet reached and exceeded the accuracy achieved by a more complex NN architecture.