Image salience object detection based on convolutional neural network

In this paper, an improved significance detection model based on FCN is proposed to solve the problem of sensitivity field limitation and computational redundancy in image significance detection of traditional full convolutional neural networks. A hierarchical network architecture including ESP module is constructed by introducing the void convolution and skip connection mechanism, and combining the multi-scale feature fusion strategy. The model adopts a progressive feature decoding scheme, including transposition convolution and channel fusion and cascade transposition convolution fusion, which can effectively solve the problem of low contrast target edge blur and suppress background noise. The experimental results show that the improved model can improve the convergence speed of cross entropy loss by 23%, the single batch inference time by 2.2%, and the pixel accuracy (PA) and average pixel accuracy (mPA) by 0.73% and 8.13%, respectively. Although the mIoU index is slightly different from the traditional method, the composite evaluation system verifies the robustness of the model in complex scenarios.