Abstract: In practical applications of compressed sensing (CS), the inability of the sampling end to access the sparsity of the original signal, combined with limited sampling resources, hinders traditional block compressed sensing (BCS) from achieving adaptive sampling rate allocation, thereby resulting in degraded reconstruction quality. To address this issue, this paper proposes an adaptive block compressed sensing method based on Support Vector Machine (SVM). The proposed method first extracts three categories of features—image block mean, variance, and local saliency—directly from the compressive measurement domain. A trained Radial Basis Function (RBF) kernel-based SVM classifier is then utilized to categorize image blocks into simple, medium, and complex classes, facilitating differentiated sampling rate allocation. In the reconstruction phase, an overcomplete dictionary combined with the Lasso algorithm is employed to further enhance the sparse representation capability of the images. Experimental results demonstrate that, compared with existing methods, the proposed approach achieves superior peak signal-to-noise ratio (PSNR) and visual quality at lower sampling rates. Furthermore, this method requires no prior information of the original image and maintains low computational complexity, making it highly valuable for image acquisition systems with constrained power consumption and bandwidth.