Few-shot image recognition represents a critical challenge in computer vision research. The scarcity of samples often results in inaccurate classification, limited generalization capabilities, and overfitted model recognition. To address these issues,...
Few-shot image recognition represents a critical challenge in computer vision research. The scarcity of samples often results in inaccurate classification, limited generalization capabilities, and overfitted model recognition. To address these issues, the present study focuses on spider image recognition utilizing transfer learning and data augmentation techniques in limited sample settings. First, the BasNet image segmentation model and background replacement algorithm are used to extract species image data from the foreground; data augmentation is then applied to address the scarcity of samples. Second, a layer-by-layer fine-tuned transfer learning strategy based on the ResNet-50 model is devised. Specifically, to mitigate overfitting in the few-shot image classification task, the first two residual blocks are frozen so that only the last two are trained. To enhance the model’s representation and generalization abilities, the SSC-ResNet-50 optimization model is constructed by introducing symmetry techniques. This study aims to enhance the accuracy and performance of spider image recognition. The experimental results demonstrate that the improved SSC-ResNet-50 model achieves an average accuracy of 99.1% in recognizing five types of spiders, thereby surpassing the performance of traditional models. These findings offer valuable insights for the field of small-sample high-precision image recognition.