Leveraging the potential of convolutional neural network (CNN) in image processing, a novel method based on GASF-CNN was introduced for detection kerosene content in gasoline. Competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA）were adopted to select the key variables, and Gram-angle and field (GASF) was used to encode the selected variables, which were then input into CNN. Experimental results revealed that using variables selected by CARS-SPA enhanced the model's performance. The root mean square error (E_RMS) of GASF-CNN on the training set and the test set was 0.620 and 0.739, respectively. The coefficient of determination (R²) on the training set and the test set was 0.988 and 0.983, respectively. However, the E_RMS on the training set and the test set of 1D-CNN, support vector regression (SVM) and partial least squares regression (PLSR) are 0.702, 0.898, 1.500, 1.290, 1.490 and 1.320, respectively; the R² on the training set and test set are 0.985, 0.975, 0.932, 0.952, 0.932 and 0.949, respectively. The amalgamation of GASF-CNN and CARS-SPA allows for more precise quantitative detection of kerosene adulteration in gasoline, thereby offering a promising methodology for spectral detection of gasoline adulteration.