EdgeNet: a low-power image recognition model based on small sample information

Abstract

Existing deep convolutional neural networks that rely on large datasets typically require images with high resolution and deep neural network models trained and called upon to improve accuracy of image recognition and classification. It is needed to use lightweight model to adapt to such low-power devices. However, lightweight small models are limited in their ability to classify and recognize small-sized images with low-resolution and are constrained by the number of parameters in the model and unable to perform deep-level feature extraction, since the low-resolution indicates small sample information. In the intelligent interaction in digital media, capturing, storing, transmitting, and computing high-resolution, high-precision images incur high power consumption and operating costs. When deploying an image recognition system on the client-side of IoT devices, it is difficult to meet the hardware requirements of high storage space and fast computation speed. It is also challenging to directly use high-resolution image data for model fine-tuning and training, and the size and parameter updates of the model are also limited by the storage and operating capacity of the hardware facilities. We proposed a low-power image recognition framework consists data pre-processing part and lightweight modeling architecture part. The data pre-processing method for image data based on an Auto-Encoder that filters R, G, B color channel data using a resolution filter to realize data compression, that is Downscaling large input data to a smaller size, thus to address the limitations of low-power deep learning model deployment and training. Based on the resolution filter, a channel normalization method is proposed to perform batch normalization on each channel dimension to encode the original image data at the same size and improve the mean squared error discrimination of the image data. And the lightweight model uses a depth-separable convolutional neural network and two kinds of blocks: one with batch normalization and the other without, EdgeNet. The architecture makes it possible to deploy more suitable for IoT device. The proposed framework achieves only a small precision loss within permission, but improves the forward inference speed of the model, and reduce the memory storage to 8.7 MB.