Although a large number of researches have been undertaken in the area of transcription start site (TSS) localization, the problem of TSS localization has not yet been fully resolved. According to the previous promoter prediction algorithm, a new sliding window based computational localization method for E. coli TSSs is proposed. The TSS-likelihood scores of each possible position in genomic sequences are calculated by the window classifier which is improved by introducing the composite motif model in the training procedure of original promoter classifier. The distribution of distances between TSSs and translation start sites (TLSs) is also utilized to calculate the TSS-position scores. Localization results are achieved from the final score profiles which combine TSS-likelihood scores and TSS-position scores. The test results on E. coli dataset show that the method can find the putative TSSs and decrease the number of false positives efficiently.