A device, computer program and method

The application is directed to improving the accuracy of the classification probability output by a neural network used in safety critical systems such as those used in autonomous and partially autonomous vehicles. An embodiment of the invention includes: a dynamic vision sensor (DVS) 115, an image sensor 120, a device 110 and a machine learning algorithm 105. The DVS 115 and the image sensor 120 are connected to the device and provide one or more images of the same field of view which are processed by the device, including compression and output to the machine learning algorithm which uses the output image to classify elements in the compressed image e.g. identification of pedestrians in the vicinity of an autonomous vehicles. The field of view of the DVS 115 and the image sensor 120 at least partially overlap and the DVS 115 may be used to determine movement of an object in the overlapping area which corresponds to a region of the image captured by the image sensor 120. The processing circuitry of the device then applies a compression algorithm to the areas of non-movement in the image received from the image sensor 120.an image from a dynamic vision sensor the algorithm applying a higher level of compression to these areas than to the areas where movement was detected which may not be compressed at all. This lower level of compression in movement areas means data loss is lower and the signal to noise ratio is higher in these areas which are those most relevant to determining a classification output by the neural network. The movement determined may be compared to a threshold with anything below the threshold being determined non-movement and therefore subject to the higher level of compression.