Homogeneous Charge Compression Ignition (HCCI) engines deliver high thermal efficiency and, therefore, low CO2 emissions, combined with low NOx and particulate emissions. However, HCCI operation is not possible at all conditions due to the inability to control the autoignition process and new understanding is required. A high swirl low-compression-ratio, optically accessed engine that can produce overall fuel lean, axially stratified charge (richer fuel mixture close to the cylinder head was achieved using port injection against open valve and homogeneous mixture during injection against closed valve timing) was operated in HCCI mode without and with spark-assist mixture ignition. The present study investigates the differences in the HCCI autoignition process and the propagation of the autoignition front with homogeneous mixture or fuel charge stratification, internal Exhaust Gas Recirculation (iEGR) (introduced by utilizing different camshafts) and spark-assisted iEGR lean combustion. In order to visualize the HCCI process, chemiluminescence flame images, phase-locked to a specific crank angle, were acquired. In addition, time-resolved images of the developing autoignition flame front were captured. Proper Orthogonal Decomposition (POD) was applied to the acquired images to investigate the temporal and spatial repeatability of the autoignition front and compare these characteristics to the considered scenarios. The eigenvalues of the POD modes provided quantitative measure of the probability of the corresponding flame structures. The first POD mode showed higher probability of single autoignition sites originating from a particular location (depending on the scenario). However, the contribution from other modes cannot be neglected, which signified multiple locations of the single autoignition and also, multiple sites of self-ignition of the fuel-air mixture. It was found that increasing iEGR resulted in random combustion (multiple autoignition sites and fronts), which, however, became significantly non-random due to addition of spark-assisted ignition. It was identified in the POD analysis of the time-resolved flame images that the presence of inhomogeneity either in the temperature or the mixture fraction distribution increases the probability of random combustion during the very early stages of flame development. Thus, the fluctuations of heat release is higher during this period.