With the popularization and application of computer vision technology in agricultural production, traditional deep learning methods have exposed limitations such as high computational cost and dependence on high-performance equipment, which restricts the transformation of agriculture to precision and intelligence. Especially in the field of crop phenotypic detection, existing methods have technical bottlenecks in conventional crop detection, and there is an urgent need to build a low-cost lightweight detection solution. To address this problem, this study proposes an ANAFD detection method based on feature optimization, which realizes efficient identification of crop fertilizer deficiency through algorithm collaborative optimization. This method constructs a three-level detection architecture: first, the improved normalized cross-correlation algorithm (NCC) is used to achieve rapid positioning of crop leaves, and the feature expression ability of the target area is enhanced by integrating Marr wavelet multi-scale feature detection and image preprocessing technology; secondly, the Adaptive and Generic Accelerated Segment Test (AGAST) + Harris Corner Detector + Non-Maximum Suppression (NMS) three-step optimization strategy is used to extract feature points, while ensuring the small-scale feature processing capability, high-quality corner points are screened through NMS. Finally, a fertilizer deficiency state determination model is established based on the morphological distribution characteristics of feature points. Compared with traditional solutions, this method has three technical advantages: (1) effectively improves detection accuracy through multi-scale feature learning; (2) reduces algorithm complexity by two orders of magnitude; (3) adapts to mainstream low-power embedded platforms. In typical crop fertilizer deficiency detection experiments, the ANAFD method has a detection accuracy of 76%–85% for crops such as cucumbers, eggplants, and tomatoes. Compared with the detection solution based on deep learning, this method eliminates complex processes such as data collection and model training, shortens the time for a single detection to 1/8 of the traditional method, reduces memory usage by 92%, and can run stably on embedded devices such as OpenMV and Raspberry Pi, providing a feasible lightweight solution for precision agriculture.