Comparison of activation function for offline handwritten Kanji document detection using convolutional neural network

In an offline kanji handwriting detection and recognition system, the ability of the neural network to correctly recognise each handwritten character within a document tends to be a significant problem. However, the present state-of-the-art neural network adopted for the object detection task settle for the object location principle but cannot achieve complete detection and lacks the proper use of an activation function. Also, there appears to be a lack of research focusing on developing an activation function that can perfectly enhance the learning ability of an artificial neuron used in a deep neural network model. Therefore, this research paper presents a visual evaluation between monotonic and non-monotonic activation function performance effect on a neural network. The results obtained show that the non-monotonic activation functions outperformed the monotonic activation function by achieving a fast speed for detection and recognition of the kanji handwritten characters.