Operating CNNs in the Presence of Faults

In the era of automation and robotics, images are captured by the camera and converted to a digital number by a CMOS sensor to carry forward the image processing by the processor. It is crucial to present the non-faulty image to the Digital signal processor(DSP) as it may not recognize it correctly. Even if the recognition is a success, there is a high possibility that DSP identifies it incorrectly because of the faults in the input image to the processor. Signal processing is mainly carried forward using deep learning algorithms. The processes that can avoid the propagation of faulty images to the DSP are manual cleaning the lens, i.e., external faults or, in case of internal fault, replacing the sensors sitting behind the camera responsible for converting light rays to the digital value i.e., CMOS sensor. But this seems a tedious process. So to solve this problem, the processors can avoid processing the faulty image, or the processor can be made reliable to deal with the faults themselves. This thesis mainly focuses on embedding faults in the image at the training of the deep learning model inside the Digital signal processor without affecting the model’s architecture, and comparative accuracy is achieved when dealing with non-faulty images or faulty images. The work is carried forward using Convolution Neural Network(CNN) that branches out from a deep learning algorithm. Furthermore, the discussion will be on the CNN model layers parameters that might get faulty during deployment and their effect on the model’s accuracy. At last, the CNN model size is compressed, resulting in-memory optimization.