Aiming at the problems of low accuracy and poor robustness in spot feature recognition under complex surveillance scenarios， a spot feature recognition method for surveillance images based on the ConvNeXt network is proposed. First， a visual saliency detection model is adopted to accurately locate the salient regions of spots in surveillance images， and segment and extract them efficiently. Then， the segmented salient regions of spots are input into the improved ConvNeXt network based on the depthwise separable convolution mechanism. Taking advantage of the unique efficient computing characteristics of this mechanism， deep extraction of multi-scale features is carried out on the salient regions of spots to fully mine the spot feature information at different scales. Finally， based on the extracted multi-scale spot features， the Softmax activation function is used to accurately identify the spot types in surveillance images， which provides a precise solution for the spot analysis of surveillance images. The experimental results show that the average quality defect index （QDI） of the method combining spot salient region segmentation with the ConvNeXt network fluctuates between 0.9 and 1.0， which is higher than that of the basic ConvNeXt network， proving that spot salient region segmentation plays a positive role in improving the performance of the recognition method. The average QDI index of the recognition method based on the ConvNeXt network fluctuates between 0.8 and 1.0， which is higher than that of the support vector machine-based recognition method， the HSV space fusion Retinex algorithm-based recognition method， and the multi-feature aggregation-based recognition method， demonstrating its high accuracy and strong robustness in the task of spot feature recognition for surveillance images.