In order to study the two-dimensional posture features of organisms with different body sizes in the context of behavioral interpretation capability, an end-to-end posture feature recognition method is proposed in this paper. The method uses ResNet-50 to extract strong image features and implement superpixel segmentation based on the image features. Based on this, the mapping relationship between superpixel features and keypoint coordinates is simulated and learned through supervised training of a fully connected network to predict the predicted keypoint location of the data input for the purpose of biological posture recognition. The results show that the method has high accuracy and robustness in gesture recognition, including 80.4% accuracy for grizzly bear keypoint detection and 94.6% accuracy for mouse keypoint detection. The results provide a theoretical basis for advanced bioinformatics understanding such as limb keypoint detection, posture 3D reconstruction, and behavior recognition.