A Single Filter CNN Performance for Basic Shape Classification

IoT cameras and sensors collect images and sensing data from everywhere in the world to transmit them via the Internet. These collected images are stacked into the servers, and an image recognition system on the server, such as CNN (Convolutional Neural Net), mines valuable information. In the near future, when the enormous number of IoTs collect images at various places, these servers would reach an overflow. Hence, if IoTs would send not only images but also analyzed results to the server, it would reduce server loads; however, the conventional CNN is too large to implement this. We propose a single-filter CNN model that can be implemented even on a small IoT. Our CNN model is of minimal configuration with an input layer, an affine transformation layer, a convolution layer, a pooling layer, and a fully connection layer. We evaluate our proposed CNN model with two experiments. First, we check whether it can learn the eleven basic shapes, i.e., a circle, a triangle, a square, etc. Second, we check whether it can classify the basic shapes against their shape reduction and their noise mixture. Results of the first experiment show that our system can classify all the basic shapes perfectly, results of the second experiment show that accuracy depends on the types of filters for both the scaled-shape classification and the inverse-pixel noise-shape classification.