Noninvasive speckle autocorrelation is a promising technique for single-shot optical imaging through scattering media. However, it fails to image multiple distinct targets within an object space through scattering media since it is constrained by the tiny effective range of optical memory effect. We present a method for multi-object single-shot imaging through scattering media that incorporates deep learning into the speckle autocorrelation technique, wherein the field of view is customized by recovered autocorrelation sidelobes, and a conventional phase-retrieval algorithm is applied to a complete set of expected speckle autocorrelations to identify multiple target objects and their relative positions. Experiments verify the feasibility of customizing the field of view for imaging through scattering media. Image reconstruction results show that the proposed approach produces superior image quality compared to existing methods. We also demonstrate its generalization capability across different object types and unknown scattering media.