Transform coding is one of the most important methods for lossy image compression. The optimum linear transform - known as Karhunen-Loeve transform (KLT) - was difficult to implement in the classic way and therefore usually replaced by the Discrete Cosinus Transform (DCT). Nowadays, due to continuous improvements of the neural network's performance, the KLT method becomes more topical then ever. In order to get compression, in all the methods based on transform coding the transform coefficients are computed and, after that, quantized. Because the quantization step is the one that makes the method lossy, it is very important for the overall performance of the method. We propose a new scheme where the quantization step is merged together with the transform step during the learning phase. The new method is tested for different levels of quantization and for different types of quantizers. Experimental results presented in the paper prove, without any doubt, that always the new proposed scheme gives better results than the state-of-the-art solution.