Application of CNN Based on Improved Activation Function in Image Classification Processing

To address the issue of weak model classification and low accuracy in image classification using Convolutional Neural Networks (CNNs), researchers have proposed a new activation function called LReLU-Softplus. This function combines the Leaky Rectified Linear Unit (LReLU) and Softplus functions. The results demonstrate that LReLU-Softplus outperforms other activation functions on four different datasets. In the self-made women's clothing image database, LReLU-Softplus achieved the highest recognition rate of 97.4%, compared to Softplus at 96.3% and ReLU at 93.2%, within the range of the study's iteration cycles. On the Mnist digit database, LReLU-Softplus had the highest recognition rate of 99.4%, surpassing Tanh at 98.7% and Sigmoid at 98.6%. In the ImageNet image library, LReLU-Softplus achieved the highest recognition rate of 81.2%, higher than LReLU at 80.6%. In the Oxford flowers image database, LReLU-Softplus achieved the highest recognition rate of 95.1%, which is also higher than the comparison functions. Overall, the proposed LReLU-Softplus activation function is effective in improving the recognition rate of image features and has high efficacy in practical CNN image classification processing.