Steganalysis as a method to detect whether image contains se-cret message, is a crucial study avoiding the imperils from abus-ing steganography. The point of steganalysis is to detect the weak embedding signals which is hardly learned by convolution-al layer and easily suppressed. In this paper, to enhance embed-ding signals, we study the insufficiencies of activation function, filters and loss function from the aspects of reduce embedding signal loss and enhance embedding signal capture ability. Adap-tive Parametric Activation Module is designed to reserve nega-tive embedding signal. For embedding signal capture ability enhancement, we add constraints on the high-pass filters to im-prove residual diversity which enables the filters extracts rich embedding signals. Besides, a loss function based on contrastive learning is applied to overcome the limitations of cross-entropy loss by maximum inter-class distance. It helps the network make a distinction between embedding signals and semantic edges. We use images from BOSSbase 1.01 and make stegos by WOW and S-UNIWARD for experiments. Compared to state-of-the-art methods, our method has a competitive performance.