This paper introduces an innovative deep joint source-channel coding (DeepJSCC) approach to image transmission over a cooperative relay channel. The relay either amplifies and forwards a scaled version of its received signal, referred to as DeepJSCC-AF, or leverages neural networks to extract relevant features about the source signal before forwarding it to the destination, which we call DeepJSCC-PF (Process-and-Forward). In the full-duplex scheme, inspired by the block Markov coding (BMC) concept, we introduce a novel block transmission strategy built upon novel vision transformer architecture. In the proposed scheme, the source transmits information in blocks, and the relay updates its knowledge about the input signal after each block and generates its own signal to be conveyed to the destination. To enhance practicality, we introduce an adaptive transmission model, which allows a single trained DeepJSCC model to adapt seamlessly to various channel qualities, making it a versatile solution. Simulation results demonstrate the superior performance of our proposed DeepJSCC compared to the state-of-the-art BPG image compression algorithm, even when operating at the maximum achievable rate of conventional decode-and-forward and compress-and-forward protocols, for both half-duplex and full-duplex relay scenarios.