2 DatasetThe Airbus Ship Detection dataset [airbus_ship_detection_2018] contains 192k satellite images, of which 22.1% contain annotated bounding boxes for a single
ship class. Key metrics of the dataset are described in
Table 1. As can be seen in the sample images in
Figure 2, a large part of the overall pixel space captures relatively homogenuous parts such as open water or clouds. We chose this dataset as it is part of the European Space Agency’s (ESA) On-Board Processing Benchmark suite for machine learning applications [obpmark], with the goal in mind to test and compare a variety of edge computing hardware platforms for the most common ML tasks related to space applications. The annotated ship bounding boxes have diagonals that vary from 1 to 380 pixels in length, and 48.3% of bounding boxes have diagonals of 40 pixels or shorter. Given that the images are 768×768px in size, this makes it a challenging dataset, as the model needs to be able to detect ships of a large variety of sizes. Since on Kaggle there are only annotations for the training set available, we used a random 80/20 split for training and validation, similarly to Huang et al [huang2020fast]. For our binary classifier, we downsized all images to 256×256px, to be compatible with the input resolution of Akida 1.0, and labeled the images as 1 if they contained at least one bounding box of any size, otherwise 0. For our detection model, we downsized all images to 640×640px in size.
RGB image size768×768Total number of images 192,555
Number of training images 154,044
Percentage of images that contain ships 22.1%
Total number of bounding boxes 81,723
Median diagonal of all bounding boxes 43.19px
Ratio of bounding box to image area 0.3%
Table 1: Summary of image and bounding box data for the Airbus Ship Detection Training dataset.
3 ModelsFor our binary classifier, we used a 866k parameter model named
AkidaNet 0.5, which is loosely inspired from MobileNet [howard2017mobilenets] with alpha = 0.5. It consists of standard convolutional, separable convolutional and linear layers, to reduce the number of parameters and to be compatible with Akida 1.0 hardware. To train the network, we used binary crossentropy loss, the Adam optimizer, a cosine decay learning rate scheduler with initial rate of 0.001 and lightweight L1 regularization on all model parameters over 10 epochs. For our detection model, we trained a YOLOv5 medium [ge2021yolox] model of 25m parameters with stochastic gradient descent, a learning rate of 0.01 and 0.9 momentum, plus blurring and contrast augmentations over 25 epochs.
4 Akida hardwareAkida by Brainchip is an advanced artificial intelligence processor inspired by the neural architecture of the human brain, designed to provide high-performance AI capabilities at the edge with exceptional energy efficiency. Version 1.0 is available for purchase in the form factor of PCIe x1 as shown in
Figure 3, and supports convolutional neural network architectures. Version 2.0 adds support for a variety of neural network types including RNNs and transformer architectures, but is currently only available in simulation. The Akida processor operates in an event-based mode for intermediate layer activations, which only performs computations for non-zero inputs, significantly reducing operation counts and allowing direct, CPU-free communication between nodes. Akida 1.0 supports flexible activation and weight quantization schemes of 1, 2, or 4 bit. Models are trained in Brainchip’s MetaTF, which is a lightweight wrapper around Tensorflow. In March 2024, Akida has also been sent to space for the first time [brainchip2024launch].
(20min delay)