This paper proposes a task-oriented semantic communication system dedicated to image data, designed to extract and transmit the information required by the receiver. The goal of the system is to recognize the semantic boundaries of the images. For this relevant challenges, a novel semantic encoder, based on compressed sensing (CS) is developed at the transmitter to extract semantic information. Additionally, a novel semantic decoder is proposed in the receiver, utilizing sparse reconstruction techniques to reconstruct semantic information. In contrast to prior studies that focused on conveying a broad spectrum of semantic information related to images along with all extracted features, this approach concentrates solely on isolating the semantic features relevant to the specific target edges. This method generates a sparse feature map, allowing for a reduction in compression rates by a novel compressed sensing technique, which implements a new sensing matrix based on polar code (SMPC). The proposed system can perform better across different signal-to-noise ratios (SNR) without requiring training data or being tied to a specific dataset. Furthermore, the study examines two distinct typical scenarios: noiseless measurements and noisy measurements. Our simulations show that with a lower compression rate, the classification accuracy, and exact recovery probability of 100 $\%$ can be attained in both scenarios.